AI News Archive: May 26, 2026 — Part 10
Sourced from 500+ daily AI sources, scored by relevance.
- Former execs of AI developer Alt found guilty of window dressing
Former execs of AI developer Alt found guilty of window dressing The Japan Times
Score: 28🌐 MovesMay 26, 2026https://www.japantimes.co.jp/news/2026/05/26/japan/crime-legal/alt-guilty-window-dressing/ - Automated Instruction Revision (AIR): A Comparison of Task Adaptation Strategies for LLM-based…
Automated Instruction Revision (AIR): A Comparison of Task Adaptation Strategies for LLM-based agents Modern LLM-based agents are becoming central to workflows that require planning and automation, tool use, and repeated decision-making, but their performance is fragile in environments that change over time. New data arriving in production often is different from the examples seen during agent development, additionally creating data and concept drift that can quickly make existing prompts outdated. The problem is intensified by model churn : organisations increasingly migrate between foundational models and versions, each with different strengths, weaknesses, and instruction-following sensitivity. In this setting, initial instructions become difficult to maintain because even small changes in the underlying model can alter the agent’s behaviour. Fine-tuning is not always the best practical solution; it demands hosting or computing budget, monitoring, and strong safeguards to avoid degrading general knowledge, safety, or alignment . It can also fragment systems into many specialised fine-tuned models across nodes or tasks, which increases operational complexity. At the same time, LLM-based agents often need continuous or periodic expert support to diagnose failures, revise instructions, and keep behaviour aligned with dynamic requirements/data. This dependency on human intervention slows adaptation and makes scaling brittle when many agents or workflows must be maintained simultaneously. The case for automated instruction revision Automated prompt and instruction revision methods offer a more structured alternative by updating agent behaviour directly from observed task performance and feedback . Such methods can reduce the need for manual re-engineering while keeping adaptation lightweight, repeatable, and easier to deploy across changing environments. They also improve transparency by making the revision process explicit, so it becomes clearer what was changed, which observations influenced the update , and which samples were excluded because they were contradictory or unsupported. For these reasons, automated instruction revision is emerging as an important direction for keeping LLM-based agents reliable, adaptable, and operationally manageable in modern agentic systems. Related work Existing approaches to automated instruction revision span a spectrum from lightweight prompt selection to full parameter adaptation. DSPy BootstrapFewShot focuses on automatically selecting high-quality in-context examples from labelled data, effectively improving performance without modifying the original instruction itself. More advanced prompt optimisation methods, such as DSPy MIPROv2 , treat instruction design as a search problem, jointly exploring candidate instructions and demonstrations to identify high-performing prompt configurations . Extending this idea, DSPy GEPA i ntroduces a reflective loop with a stronger teacher model that iteratively revises instructions based on execution feedback, enabling more informed and adaptive improvements. Finally, TextGrad frames instruction revision as a continuous optimisation process in text space, updating the prompt through gradient-like signals derived from task loss while keeping the underlying model fixed. Together, these methods illustrate different trade-offs between simplicity, interpretability, computational cost, and adaptability in automated instruction revision. There are other methods like CoachLM and Prewrite , but to sum up, the literature reveals two main tendencies. The first is the movement from manual prompt writing toward automated optimisation of prompts, examples, and full LLM pipelines. The second is the growing interest in interpretable decision structures such as rules or trees. Our work lies at the intersection of these directions: like prompt optimisation methods, it seeks to automatically improve LLM behaviour , but like tree-based approaches, it emphasises the discovery of explicit and reusable decision logic. What AIR does differently Our proposed Automated Instruction Revision (AIR) method is the same data-driven prompt adaptation pipeline that learns explicit task guidance from labelled examples instead of relying on continuous manual prompt engineering. But unlike the described methods, the central idea is to transform supervised task data into a compact set of task-specific decision rules, convert these rules into an executable system prompt, and iteratively refine them using additional sampled cases from the training set. This approach helps to reach comparable/top results with computational budgets that are times lower . If you’re interested in achieved results, please proceed to the RESULTS section directly. The AIR pipeline The AIR pipeline proceeds through clustering, rule induction, rule aggregation, and targeted refinement. Starting from labelled training data, it constructs an executable instruction set in the following stages. So the very first task is to create your init prompt, which can be pretty simple, like “ Please, classify/answer according to some primary rule/instruction… ”. In our benchmarking, we call this the initial prompt performance. Then we need to process our dataset: as stated above, here we need some feedback or supervision, thus it is not a self-learning approach, similar to comparison improvement techniques. So here AIR first maps each dataset into canonical input and output columns and computes embeddings for both sides (independent and target) of the supervision signal. This creates a uniform representation across tasks and provides the structure used later for clustering and local comparison. Then we should understand how to sample our data and form batches. This task is important because we can’t overload context with the entire dataset (even if it fits 1M tokens or any future RoPE -like improved size). Also, when we feed, for example, some sorted or one-class or fully randomly formed batch, it won’t help us efficiently to form any semantic decision boundaries. Our idea is to make some initial semantic grouping of the inputs, but take different targets (or mimic the targets' distribution in the general sample). This can be achieved by clustering using a few objectives (input similarity, target variety). The number of clusters is treated as a hyperparameter and is set to the default value of 5 in our experiments. Finally, AIR performs a repair step (automated blind clustering is not always perfect for different tasks (like discrete classes vs enhanced answers)) that redistributes samples from clusters containing only one output class to nearby alternative clusters. This encourages groups to remain semantically coherent while still preserving output variation that is useful for distinguishing behaviour. Then comes the rule learning part, which consists of a few steps. Firstly, after sampling balanced batches, these batches are passed to a reasoning model with an instruction to infer a small number of rules of the form “ if condition on the input, then output action or pattern. ” The purpose is not free-form explanation, but extraction of compact decision-boundary rules that distinguish competing output behaviours within the same semantic neighbourhood. After this, we need to aggregate and compile the rules into an executable prompt: Because rule induction is repeated across clusters, the initial rule pool is usually large and partially redundant. AIR therefore applies a dedicated LLM-based rule-compilation stage. In this stage, a reasoning model receives the induced rules and, following a constrained compiler prompt, groups rules with semantically similar THEN actions, identifies the shared structure of their IF conditions, merges complementary signals, removes lexical noise, and preserves only the exclusions needed to avoid cross-rule collisions. Aggregated rules produced by the above step are then concatenated with the task description to form a structured final prompt that instructs the model to follow the decision process encoded in the rules. AIR also creates a traced variant of this prompt that asks the model to return both the final prediction and the identifiers of the applied rules, enabling later refinement. Talking about refinement , this also appeared in several of the mentioned approaches, and it is almost about the continuous improvement process. Here, rather than using the aggregated rules as final, AIR re-evaluates them on newly sampled examples and records predicted outputs, applied rule identifiers, and task metric values. For each rule, the pipeline separates participating cases into mistakes, where the prediction was incorrect, and anchors, where the prediction was correct according to the provided metric. Small refinement batches built from these two sets are then sent back to the reasoning model, which is asked to make the smallest necessary local revision to the current rule. In this way, AIR updates individual rules while attempting to preserve behaviour on successful anchor cases: After refinement, the updated rules are assembled into the final system prompt used for downstream evaluation. A very important note here is that the final step will be used as the major part of the agent's continuous improvement loop , especially without the involvement of AI specialists . In this way, AIR derives task-specific instructions from labelled examples through rule induction, aggregation, and revision: Performance and benchmarks This section describes the benchmark suite, the compared adaptation methods, and the models used in our experiments. The evaluation is designed to support a controlled comparison of prompting, retrieval-based adaptation, prompt optimisation, and fine-tuning under a shared task-specific setup. We evaluate AIR on a deliberately diverse benchmark suite chosen to reflect different sources of adaptation difficulty rather than a single task family. The suite spans remapped-label classification, closed-book factual QA, schema-constrained extraction, PII identification, and event-order reasoning. The comparison is made via classical cross-validation using train, development, and test splits, with task-specific metrics (typically, the higher the better ). The test set, which was outside the training and refining/tuning, was used for the final evaluation and is present in our resulting benchmarking tables. Let's then review our tasks and datasets, which partially appear in different benchmarks and can be used for our evaluation. The first one is the example of a classification task , which typically can be solved by many different and proven simpler ML-based approaches rather than using LLMs (at least to decrease complexity/computing and increase the consistency and transparency), but sometimes the dependencies can be very close to huge variety of different semantics/knowledge behind, where the LLM-based agent with structured output may be the right way. Here, we used real customer support requests collected from Twitter, based on the Customer Support dataset (sample example on Kaggle ), and constructed an 8-class benchmark by selecting requests addressed to eight companies. The main problem was that this is not a complicated task for LLM, which has a background at least around the companies that appeared in the dataset (or rather has seen the dataset during training/caught keyinfo distillation, etc.). Thus, firstly, we’ve removed any explicit company mentions and other direct lexical cues and reassigned the output labels so that each company is consistently mapped to a different description rather than its original one. The final dataset is balanced across the selected companies, and performance is measured with exact match (accuracy). Another, less discrete task is closed-book question answering , where we built a benchmark from Ever Young by Alice Gerstenberg and created question-answer pairs that must be answered without providing the full source text at inference time. The benchmark is intentionally designed so that the relevant source material is unlikely to be part of the model’s pretrained knowledge, but such leakage is valid for all LLM-based approaches, which we are benchmarking and is not decisive, which you’ll see from the evaluation. To evaluate outputs, we use an LLM-as-a-judge rubric with four dimensions: correctness, completeness, absence of hallucination, and focus, each scored on a 0/0.5/1 scale (each interval has described decision boundaries, the same as for other tasks below, where there are such stochastic native measurements), with the final score defined as the average of the four subscores. For information retrieval , we use campaign-finance filings for Philadelphia elections from the City of Philadelphia Campaign Finance Reports source, containing transaction-level records for contributions, expenditures, and debt. Each example is transformed from a structured table row into a raw CSV-style string consisting of a header and one data line, and the columns are randomly shuffled. This means the model must first reconstruct the field mapping before it can find and extract the requested outputs (this is not the ETL-related or data extraction tasks). In addition to recovering directly stated fields, the model must infer two derived labels: candidate_relationship, which categorises how the donor relates to the candidate, and contribution_context, which combines donor type and contribution size into labels. Performance is measured as average exact match across the two target fields (approx MAP). As we stated, we’ve tried to add different tasks and benchmarks to provide an overview of existing and proposed improvement approaches from as many possible perspectives. And the next task is PII extraction . Here, we used PUPA (Private User Prompt Annotations), a benchmark derived from real WildChat user-assistant conversations that contain explicit PII leakage. In our benchmark, we isolated only the annotation component and converted it into a pure extraction task: given user_query, the model must output the gold pii_units directly, without rewriting, redaction, or response generation. Evaluation is performed with an entity-level exact-match F1 score. For event logical reasoning , we adopt the Event Logic Reasoning subset of BizFinBench.v2 . Each example describes a finance-related scenario involving multiple market events, and the task is to recover their correct logical order, either temporally or through cause-and-effect structure. The benchmark is intended to test whether this ordering logic can be induced through prompting, transferred from similar retrieved examples, or absorbed through fine-tuning. The output is a compact event-index sequence, for example, “2,1,4,3,” rather than a free-form explanation. Performance is measured with an exact match. We compare AIR against a set of adaptation strategies (partially listed at the beginning as related approaches) implemented within a shared task-specific workflow across our benchmarks. All methods start from the same manually written task instruction and use the available labelled training data, but they differ in whether they adapt the prompt, retrieve examples, or update model parameters. Initial prompt baseline. The model is evaluated with the manually written task prompt only, without retrieval, optimisation, or parameter updates. This baseline measures the performance of direct zero-shot prompting with human-authored instructions. KNN-based prompting . For each test example, we retrieve (this step is similar to the corresponding one in the RAG pipelines , which nowadays is probably more known for non-DS/AI/ML specialists and business), similar training samples, and append them as in-context examples to the initial prompt. This baseline tests whether instance-level adaptation through dynamic example selection is sufficient without modifying the instruction itself. DSPy BootstrapFewShot. DSPy automatically compiles a few-shot prompt by sampling labelled training examples and retaining demonstrations that satisfy the task metric. This yields an automatically selected demonstration set while keeping the base instruction fixed. Fine-tuning. We fine-tuned the base underlying foundation model (GPT family, as proposed in other improvement papers — described below after this section) on the training datasets using chat-formatted supervision constructed from the same initial system prompt, the task input, and the target, and the resulting model is evaluated on the test dataset. This baseline represents parameter adaptation rather than prompt-only adaptation. DSPy MIPROv2 . MIPROv2 is used as a prompt optimisation baseline without a separate teacher or reflection model. In our experiments, it is run with DSPy’s auto=”medium” budget and task-specific limits on bootstrapped and labelled demonstrations, which are chosen separately for each benchmark. It jointly searches over instruction candidates and, when enabled, demonstration candidates to find a high-performing compiled prompt. DSPy GEPA . GEPA is used as a reflective prompt optimiser with the same auto=”medium” budget, but unlike MIPROv2, it relies on a separate reasoning model as a teacher for reflection (in AIR, we called the reflection conducted in a different way as refinement). So, the GEPA algorithm receives task-specific textual feedback through the metric and uses the teacher model to revise candidate instructions based on execution traces and validation-time search. TextGrad . TextGrad treats the system prompt as an optimizable text variable while keeping the model parameters fixed. In our experiments, we run one epoch over the full training set; within each minibatch, per-example losses are summed into a single minibatch loss, which is then used to update the prompt. This baseline, therefore, represents iterative text-space optimisation of the initial instruction. AIR . AIR induces explicit task-specific rules from labelled examples, aggregates them into a compact rule set, and refines them iteratively before assembling the final prompt. Across this benchmarking, we used gpt-4.1-mini-2025–04–14 as the base task model for evaluation, at least because the GPT family was proposed and used in the described existing methods. This model is also used for the initial prompt baseline, KNN-based prompting, DSPy BootstrapFewShot, MIPROv2, and the final prompt evaluation in AIR, and it also serves as the starting point for fine-tuning. For methods that require a stronger auxiliary model for reflection or instruction refining/revision, we use gpt-5–2025–08–07 as the teacher model in GEPA, TextGrad, and AIR. In AIR, we additionally use text-embedding-3-small to compute input and output embeddings for clustering and rule induction. For the closed-book question answering benchmark, the LLM-as-a-judge metric uses gpt-5.1 as the judge model. Results This section reports the quantitative results for all compared methods across the five benchmark tasks. In addition to the task metric, each table includes the token usage reported in the original experiment slides, separating training-time token usage by model role and inference-time token usage for the base model. Below is the first table showing the classification results . GEPA achieves the highest score, while AIR remains very close and also surpasses fine-tuning. This pattern suggests that, once brand priors are removed, the task behaves less like ordinary text classification and more like learning a remapped latent label system, where explicit prompt-level instruction discovery can be highly effective. Please also compare the computational budgets (or tokens used), which are quite high for DSPy/TextGrad families. The closed-book QA results in Table 2 show different problem settings. KNN clearly performs best, while most prompt-optimisation methods fail to improve over the initial prompt. This shows that the benchmark is dominated by source-specific knowledge injection rather than by generic reasoning or procedural instruction improvement. Table 3 shows that fine-tuning is dominant on the information retrieval task: This is consistent with the structure of the benchmark: before performing extraction, the model must first reconstruct the field mapping from shuffled CSV-style rows. AIR performs substantially worse here (the same as other instruction revision methods), due to the difficulty of capturing this task through compact local rules alone. The PUPA results in Table 4 again favour fine-tuning, with MIPROv2, AIR, and GEPA forming a second tier. This pattern suggests that the benchmark rewards not only general PII detection, but also adherence to dataset-specific annotation habits. AIR remains competitive with GEPA, but does not match the best parameter-adaptation result. Note: In the source results, AIR on PUPA was terminated after 8 steps, including 5 consecutive steps with no improvement, using a batch size of 3. Finally, Table 5 shows that fine-tuning again performs best on event logical reasoning . The relatively strong initial-prompt score suggests that the base model already has some latent capability for financial event ordering, while fine-tuning appears to stabilise how that reasoning is mapped into the required output sequence. KNN and AIR both provide moderate gains, but neither approaches the fine-tuned model: Conclusions AIR is intended to provide several practical advantages . First, it is interpretable: the learned adaptation is represented as readable instruction text rather than hidden inside parameter updates. Second, it can support rule-level inspection and revision: individual rules can be examined and modified without changing model weights. Third, the pipeline preserves intermediate artefacts that can help analyse how the final prompt was constructed. Fourth, it reduces manual effort by automating a process that would otherwise require a person to inspect the data, identify recurring patterns, and summarise them as explicit rules. Fifth, the computing budget is quite moderate in comparison to other methods. At the same time, AIR has important limitations . Its core assumption is that useful task behaviour can be guided by natural language or semantic rules. When labels are inconsistent, examples are noisy, or correct decisions depend on latent patterns that are difficult to verbalise, rule induction may become unstable. AIR can also suffer from rule interaction effects: a locally beneficial instruction may conflict with others after aggregation, and repeated revisions may reduce clarity instead of improving it. The above benchmarking results do not support a single universal winner across all benchmarks. Instead, they show that different methods are strongest under different task conditions. Retrieval works best when the task is dominated by source-specific knowledge, fine-tuning is strongest when the task depends on stable dataset-specific mappings or output conventions, and instruction-based adaptation is most competitive when the target behaviour can be expressed as explicit decision rules. Within this picture, AIR occupies a clear practical niche. It does not win on every benchmark, but it remains especially competitive on tasks where the target behaviour can be captured as reusable rules within the optimised computational budget. This makes AIR attractive in settings where interpretability is important, and some loss relative to the best task-specific method is acceptable. Another important point is that the compared adaptation methods still rely on a separate search or optimisation stage to identify a strong task-specific solution. When the data changes, this process may need to be repeated rather than updated incrementally. In principle, AIR may offer an advantage here, since new data could potentially be incorporated by extending or refining the rule set instead of rebuilding the adaptation from scratch. However, this possibility was not evaluated in the present experiments and remains a direction for future work. For these reasons, we treat AIR not as a universal solution to task adaptation, but as a structured and interpretable option for settings in which task behaviour can be captured, at least in part, by explicit instruction rules derived from supervision. Building LLM-based agents that need to keep up with changing data? Talk with our experts . Read more about our data science work on our blog Connect with us on LinkedIn Do you have questions about the method or the benchmarks? Feel free to ask in the comments. This article is an adapted version of the original research paper, edited for Medium. The full paper is available at https://arxiv.org/abs/2604.09418 Automated Instruction Revision (AIR): A Comparison of Task Adaptation Strategies for LLM-based… was originally published in DataDrivenInvestor on Medium, where people are continuing the conversation by highlighting and responding to this story.
- When It Comes to Your Personal Brand, Don't Let AI Flatten Your Voice
“Editing your LLM output doesn't have to be any different than how your English teacher hurt you with that red pen, said Tanya Svoboda, senior content manager at Workday. "You are not a prompt.
- Grow Predictably Launches Diagnose-First AI Adoption Consulting for B2B Marketing Leaders
Grow Predictably Launches Diagnose-First AI Adoption Consulting for B2B Marketing Leaders azcentral.com and The Arizona Republic
- The real risk of AI is marketing commoditization
AI can speed up every marketing team. The bigger challenge is avoiding commoditization when everyone uses the same tools. The post The real risk of AI is marketing commoditization appeared first on MarTech .
- Gen Z is the AI generation. So why are they calling travel advisors more than Boomers?
Gen Z is the AI generation. So why are they calling travel advisors more than Boomers? Raleigh News & Observer
- How the Impact Analysis Agent ROVO helps you handle your Jira tickets smarter
How the Impact Analysis Agent ROVO helps you handle your Jira tickets smarter Atlassian Community
- Stop Using LLMs Like Giant Problem Solvers
How I turned 100 messy pdfs into structured insights by building a deterministic loop around agents The post Stop Using LLMs Like Giant Problem Solvers appeared first on Towards Data Science .
Score: 27🌐 MovesMay 26, 2026https://towardsdatascience.com/stop-using-llms-like-giant-problem-solvers/ - When AI evolves on its own (KOR)
Kim Dae-shik The author is a professor at KAIST. Ancient Greeks used the word "mythos" to describe stories or narratives. Like all words, its meaning shifted over time. Beginning in the early 19th century, Europeans increasingly applied mythos only to old tales, especially ancient Greek legends. That is why the word today is often understood simply as “myth.” Anthropic's chief product officer, Ami Vora, its co-founder and president, Daniela Amodei, and co-founder and CEO Dario Amodei present on stage at the Code with Claude developer conference in San Francisco on May 6. [AP/YONHAP] Yet another interpretation exists. In “Poetics” (335 B.C.), Aristotle used mythos to describe the representation or structure of action within tragedy. By contrast, ethos referred to the character performing the action, while praxis meant the action itself. Aristotle argued that the essence of tragedy lay not in character or action alone but in mythos, which connected the two. OpenAI CEO Sam Altman, center, and Greg Brockman, OpenAI president and co-founder, arrive at the federal courthouse during proceedings in a lawsuit against OpenAI in Oakland, California, on April 30. [AFP/YONHAP] That background makes the naming of the latest AI model introduced by Anthropic on April 7 symbolic. The important point is not who created it or how it was built, but the very existence of an AI system with such capabilities. Mythos is described as the most advanced AI system developed so far, especially in coding and hacking. Reports say it identified vulnerabilities not only in operating systems such as Windows and macOS but also in BSD Unix, which has supported much of the global internet infrastructure for nearly three decades. It reportedly went further by proposing hacking strategies based on those vulnerabilities. In theory, if such technology were obtained by terrorist groups or rogue states, it could threaten the global internet system. Online payments, logistics systems and communications infrastructure could all be affected simultaneously. Such dystopian possibilities no longer seem entirely unimaginable. As a result, Anthropic reportedly decided not to release Mythos publicly. Instead, access has been limited to companies connected to the discovered vulnerabilities. That decision raises another issue. Most companies participating in the so-called Project Glasswing are U.S. firms. Korean companies, meanwhile, reportedly cannot access Mythos. The imbalance means that while the U.S. government and certain private companies may now possess tools capable of disrupting foreign information technology infrastructure, many countries, including Korea, lack comparable defensive capabilities. Related Article Rise of AI raises fears of North Korean hacking capabilities The butterfly effect of the Anthropic contract termination OpenAI officials discuss safety protocols with Canada following school shooting OpenAI claims China's DeepSeek trained its AI by distilling U.S. models, memo shows Another concern lies in the nature of modern AI competition. Today’s AI race is driven less by different theories or algorithms than by scale. Since most companies rely on similar mathematical foundations, the decisive factors are massive datasets and large-scale GPU infrastructure. Given enough computing power, creating comparable AI systems eventually becomes a matter of time. Reports suggest that GPT-5.5, released by OpenAI on April 23, possesses coding and hacking abilities similar to Mythos. In effect, Mythos may mark the beginning of unlimited competition among major technology companies. For U.S. allies such as Korea, this competition presents both opportunities and risks. For China, however, it represents a strategic vulnerability. The obvious response for Beijing would be to develop an AI system comparable to Mythos. Yet China’s flagship AI model, DeepSeek V4, released on April 24, reportedly has not demonstrated the performance many expected. Because of semiconductor export controls, China still faces difficulty building data center infrastructure on the scale available in the United States. Under such conditions, matching AI models developed by Anthropic, Google or OpenAI remains challenging. If China cannot catch up through conventional methods, it may seek alternatives. One option would be integrating multiple AI systems developed by different firms into a single national AI champion. Another could involve nationalizing domestic data centers to create state-led computing infrastructure. Japanese Finance Minister Satsuki Katayama speaks to the media after a meeting involving the Financial Services Agency, the Bank of Japan, the National Cybersecurity Office, the country's top three banks and the Japan Exchange Group following concerns about potential vulnerabilities linked to Anthropic's Mythos AI model in Tokyo on April 21. [REUTERS/YONHAP] But if even those measures fail, China could turn to a more dangerous path: recursive self-improvement, or RSI. Proposed in 1965 by British mathematician I. J. Good, RSI refers to a process in which AI rewrites its own code to improve its intellectual capabilities. If successful, such systems could rapidly evolve into artificial superintelligence. During the Cold War between the United States and the Soviet Union, the doctrine of mutually assured destruction, or MAD, rested on the paradox that the ability to annihilate each other deterred the use of nuclear weapons. The article argues that AI, accelerated by systems such as Mythos, is beginning to transform into a strategic weapon surpassing nuclear arms. Unlike the nuclear rivalry of the 20th century, however, the ultimate winner in a 21st-century AI version of MAD between the United States and China may be neither country. Instead, it could be the artificial superintelligence created through recursive self-improvement itself. 인공지능이 스스로 진화할 때 김대식 KAIST 교수 고대 그리스인들은 이야기나 스토리를 미토스(mythos)라고 불렀다. 하지만 어차피 모든 단어는 시간이 지나면 의미가 변하는 것이지 않았던가. 19세기 초부터 유럽인들은 미토스를 오래된 이야기, 특히 고대 그리스 전설에만 적용하기 시작한다. 오늘날 미토스를 우리가 전설로 이해하는 이유겠다. 하지만, 미토스에 대한 다른 해석도 가능하다. 바로 아리스토텔레스가 시학에서 제안한 비극에서의 역할이다. 아리스토텔레스는 미토스를 비극에서 벌어지는 행위의 재현이나 묘사로 해석한다. 반대로 에토스(ethos)는 행위를 하는 인물, 그리고 프락시스(praxis)는 행위 그 자체를 의미한다. 그리고 비극의 핵심을 아리스토텔레스는 인물이나 행위 그 자체보다는, 인물과 행위를 연결해주는 바로 미토스라고 제안했다. 지난 4월7일 소개된 가장 최신 AI 모델의 이름을 미국 엔트로픽 사가 미토스라고 정한 것은 대단히 상징적이다. 누가, 어떤 식으로 만들었는지는 별로 중요하지 않다. 미토스라는 인공지능이 등장했다는 사실 자체가 가장 중요하다. 지금까지 만들어진 인공지능 중 가장 뛰어나다는 미토스. 특히 코딩과 해킹 능력은 압도적으로 최고라고 한다. 덕분에 미토스는 윈도우나 맥OS같은 컴퓨터 운영체제뿐만이 아닌, 지난 30년 가까이 글로벌 인터넷망을 뒷받침하던 BSD 유닉스의 버그까지도 찾아냈고, 그런 버그들을 기반으로 해킹 전략까지도 제안했다고 한다. 테러단이나 불량국가 손에 들어가는 순간 이론적으로는 글로벌 인터넷망이 무너질 수도 있다는 말이다. 단순히 이메일이나 SNS를 넘어 온라인 결제, 물류, 통신 인프라가 하루아침에 무너지는 디스토피아 적인 상상도 더는 불가능해 보이지 않는다. 덕분에 엔트로픽은 미토스를 출시하지 않기로 결정했고, 대신 버그가 발견된 기업들에만 한정적으로 제공하기 시작했다. 그런데 여기서 문제가 생긴다. ‘프로젝트 글래스윙’이라는 이름으로 시작된 이 프로젝트 멤버들은 대부분 미국 기업이다. 우리 한국 기업들은 여전히 미토스 사용이 불가능하다. 미국 정부와 특정 기업들은 이제 타 국가 IT 인프라를 무너트릴 수 있는 능력을 갖추게 되었지만, 대한민국을 포함한 대부분 국가는 방어할 수 있는 능력이 없다. 그리고 또 하나의 문제가 있다. 현대 인공지능의 핵심은 알고리즘이나 이론보다 규모의 경쟁이다. 어차피 모두 비슷한 수학적 이론과 알고리즘을 활용하기에, 인공지능 학습에 필요한 데이터, 그리고 빅테크 스케일의 GPU만 가지고 있다면, 비슷한 수준의 AI를 만드는 것은 시간문제다. 실질적으로 4월23일 오픈AI가 출시한 GPT-5.5는 코딩과 해킹 면에서 미토스와 비슷한 능력을 갖추고 있다고 보고되고 있다. 미토스를 시작으로 빅테크들 간의 무한경쟁이 시작된 것이다. 이렇게 미국에서 시작된 빅테크들 간의 경쟁은 대한민국 같은 동맹국에는 기회와 리스크가 동시에 존재하는 양면의 칼이겠지만, 중국에는 치명적인 전력적 리스크일 수밖에 없다. 그렇다면 답은 하나다. 중국 역시 자체적으로 미토스 수준의 인공지능을 개발해야 한다. 하지만 지난 4월24일 출시된 중국 대표 AI 모델인 딥시크 V4는 기대했던 만큼의 성능을 보여주지 못하고 있다. 반도체 수출 규제 때문에 여전히 미국 수준의 데이터 센터 인프라를 구축하지 못한 중국에서 앤트로픽·구글·오픈AI 수준의 인공지능 모델이 출시되는 것은 여전히 어려워 보인다. 기존 방법으로 미국을 따라잡을 수 없다면, 중국은 새로운 방법을 택할 수도 있다. 우선 여러 기업이 개발하고 있는 AI 모델들을 하나로 통합해 국가 AI 챔피언을 키워볼 수 있고, 중국 내 데이터 센터들을 국영화해 범국가적 데이터 센터를 구축해 볼 수도 있겠다. 하지만 만약 그런 방식을 사용해도 미국 수준 AI 모델을 만들 수 없다면, 중국은 어쩌면 가장 위험한 방법을 선택할 수 있다. 바로 AI의 ‘재귀적 자기 개선’이다 (Recursive self-improvement, RSI). 1965년 영국 수학자 어빙 구드가 제안한 RSI에서는 인공지능이 자신의 코드를 스스로 재작성함으로써, 자신의 능력과 지적 역량을 향상시키는 과정을 의미한다. RSI가 성공한다면, AI의 지능은 폭발적으로 향상되어 초지능 (Artificial Super intelligence, ASI) 수준까지도 도달할 수 있다. 20세기 미국과 구소련 사이 냉전 시대에는 ‘상호확증파괴(MAD·Mutually Assured Destruction)’이라는 전략이 있었다. 언제든지 서로를 전멸시킬 수 있다는 가능성 그 자체가 핵무기 사용을 억제한다는 역설적인 전략이었다. 미토스 덕분에 핵무기를 능가하는 새로운 전략적 무기로 변신하기 시작한 인공지능. 하지만 20세기 핵무기 경쟁과는 달리 21세기 미국과 중국 사이 MAD 전략의 승자는 결과적으로 미국도 중국도 아닌 재귀적 자기 개선을 통해 만들어질 초지능, ASI일 수도 있다. This article was originally written in Korean and translated by a bilingual reporter with the help of generative AI tools. It was then edited by a native English-speaking editor. All AI-assisted translations are reviewed and refined by our newsroom.
- Beyond automation: Here’s how the power of AI can really drive your organization forward
The key to achieving those results is tapping into the data analytic superpower of AI.
Score: 26🌐 MovesMay 26, 2026https://www.hrdive.com/spons/beyond-automation-heres-how-the-power-of-ai-can-really-drive-your-organiz/820815/ - Embark Marketing Shares How Restaurants Are Actually Using AI in 2026
Embark Marketing Shares How Restaurants Are Actually Using AI in 2026 azcentral.com and The Arizona Republic
- ‘We can stitch together our past’: the AI-generated time-travellers vlogging from history
The content creators behind channels like Chloe VS History are using AI tools to ‘bring history to life in a really visceral way’ “I have just arrived in Tudor London, 1536,” a young woman in a green puffer jacket tells the camera. “I’m going to check in at my room in the inn, get into the market. Then, later I am meeting the actual king – yep, Henry VIII – in person.” On YouTube and other social platforms, users are flocking to watch AI-generated “history influencers”, characters that vlog their travels to historical settings. Continue reading...
- SalesCloser Secures Second U.S. Patent for AI-Driven Call Automation Technology
SalesCloser Secures Second U.S. Patent for AI-Driven Call Automation Technology Toronto Star
- Jarvis: Parenting teens in the age of AI means choosing trust over control
Jarvis: Parenting teens in the age of AI means choosing trust over control The Mercury News
- Dow Jones Futures Rise After S&P 500, Nasdaq Hit Highs; 5 AI Stocks In Buy Areas
The S&P 500, Nasdaq and Russell 2000 hit new highs as Micron surged to top a $1 trillion market cap. Caterpillar, Viavi flashed buy signals. The post Dow Jones Futures Rise After S&P 500, Nasdaq Hit Highs; 5 AI Stocks In Buy Areas appeared first on Investor's Business Daily .
Score: 25🌐 MovesMay 26, 2026https://www.investors.com/market-trend/stock-market-today/dow-jones-futures-sp-500-nasdaq-micron-1-trillion/ - I used Gemini in Google Maps to pick my next pizza spot — and the top result totally surprised me
I used Gemini in Google Maps to pick my next pizza spot — and the top result totally surprised me Tom's Guide
- After years of testing robot mowers, I found the specs you can safely ignore (and which ones truly matter)
Robot mowers aren't one-size-fits-all devices - many have overhyped specs you don't even need. Here's what I recommend skipping.
Score: 25🌐 MovesMay 26, 2026https://www.zdnet.com/article/robot-mower-buying-guide-features-to-ignore-2026/ - Bifrost pitches 3-D AI training platform to Korean manufacturers
Charles Wong, co-founder and CEO of San Francisco-based startup Bifrost, speaks during an interview with the Korea JoongAng Daily at Coex in Gangnam District, southern Seoul, on May 20. [AWS] [INTERVIEW] Bifrost, a San Francisco-based startup, leverages computer vision and 3-D generation technologies to train so-called physical AI systems, a service that co-founder and CEO Charles Wong believes could meet the growing needs of Korean manufacturers and tech companies. "There's a huge manufacturing base in Korea, and you build your own products ranging from semiconductors, electronics and motors — there's a nice big ecosystem of all these pieces coming together, which not many other countries have strong points in all of them," said Charles Wong, co-founder and CEO of Bifrost, in a recent interview with the Korea JoongAng Daily during the AWS Summit Seoul 2026, held between Wednesday and Thursday at Coex in Gangnam District, southern Seoul. Bifrost has recently been eyeing expanding into the Korean market. Its current sole Korean customer is an autonomous maritime startup, Seadronix, to generate the synthetic training and testing data that the company uses to develop its autonomous ship navigation AI. Bifrost is now building a similar platform centered on robotics, a push that has accelerated since Nvidia CEO Jensen Huang declared last year that physical AI and robotics represent the next major frontier. Wong sees Korea as the ideal market for that expansion. "We just completed a successful trial with a major Korean company for home robotics use cases," he said. "This company is doing development for future generations of robotics products, and they want to use our tools to speed up development and enable new capabilities." Beyond that trial, Bifrost is in talks with several Korean conglomerates about partnerships, which Wong said he looks forward to announcing when they are finalized. Wong co-founded Bifrost in 2020 after working on AI perception models for self-driving cars at NuTonomy, an MIT spinout that ran the world's first autonomous taxi trial in Singapore in 2016. Even then, before AI had become part of daily life, he was convinced that autonomous technology would eventually reach every industry — from homes and factories to ships and aircraft. Charles Wong, co-founder and CEO of San Francisco-based startup Bifrost, speaks during an interview with the Korea JoongAng Daily at southern Seoul's Coex on May 20. [AWS] "It was just a matter of when," he said. "So we thought about what the bottleneck was stopping AI from coming into the real world, and that was data. We had to find a better way to create huge amounts of organized data for physical AI to learn from, and that's how we landed on synthetic data." Bifrost's core product is Stardust, a synthetic data platform that allows an AI developer with no simulation experience and no dedicated 3-D team to produce high-quality simulated environments for training AI models. Bifrost's goal is not to replace real-world data entirely, but to give developers a way to identify weaknesses in their AI systems early and cheaply — so by the time they move to expensive real-world testing, they already know exactly what to look for. "A developer can go from zero to a ready dataset in about five minutes," Wong said. Despite having a team of around 30 employees, Bifrost's customers span multiple sectors — from maritime and aerial to defense and government — including the U.S. Air Force, NASA and defense technology firm Anduril. The startup has raised a total of $13.7 million across four funding rounds, with the latest $8 million secured from Carbide Ventures, Airbus Ventures, Peak XV's Surge and several more. A screen capture of Bifrost website [BIFROST] Underpinning all of this is Bifrost's deep reliance on AWS infrastructure, which provides the computing power, storage and AI tooling that allow the company's simulation software to operate at scale. "AWS has some great features that are really important to us," Wong said. "SageMaker lets us run inference workloads very quickly, which means our customers can validate their perception systems — basically test whether their AI can actually see and understand the world correctly — in a very short time. "The other big one is Bedrock. Right now, our users write Python code to build scenarios, but a lot of them are already using tools like Claude Code or Copilot to do that very quickly by connecting those tools to our documentation. It's very natural that at some point we bring that functionality into the app itself, through something like Bedrock — so instead of writing code by hand, users can just describe what they want and an AI agent builds the scenario for them." BY LEE JAE-LIM [lee.jaelim@joongang.co.kr]
- This $40 AI tool gives crypto traders five years of real-time market signals
Invest wiser with this five-year subscription to Biyond AI-Powered Crypto Analyst.
Score: 25🌐 MovesMay 26, 2026https://mashable.com/tech/may-26-biyond-ai-powered-crypto-analyst-basic-plan - How I Built a Real-Time In-Car SOS Detection System With Qdrant Edge, SigNoz, and YAMNet
The Button Nobody Presses Think about the last time you drove through a sketchy stretch of highway at night. Or the time you had a medical episode in the car and had to focus on not crashing the vehicle. The classic response to an emergency is a button. A physical SOS button, or some voice-activated assistant, or a crash detection feature buried three menu levels deep in an app. All of them share the same flaw: they require you to do something. They require you to be conscious, coherent, and aware enough to reach for help. That is a remarkably bad assumption in an emergency. What I wanted to build was something different. A system that listens passively, all the time, entirely on the device, and sends an alert automatically the moment it detects that something is wrong. No buttons or commands. No cloud uploads of your raw data. One major challenge was ensuring the entire detection pipeline stayed within the system’s 500ms processing window. Every 500ms, the device captures fresh audio, generates a YAMNet embedding, and runs a Qdrant Edge similarity search so if processing takes longer than that, the pipeline starts buffering and eventually breaks down. Using OpenTelemetry with SigNoz , I traced each stage of the workflow and found that YAMNet embedding generation took ~22ms while Qdrant Edge search stayed under 1ms. This gave real visibility into performance bottlenecks without ever sending raw audio off-device. The full implementation is on GitHub . You can follow along with the code as you read through this. What We Used and Why? — Qdrant Edge The standard approach to vector search involves spinning up a cloud database (like Pinecone or the standard Qdrant) over HTTP. For cloud apps, this is fine. For an in-car edge device , it is completely untenable. You cannot depend on cellular networks to save a life, you cannot run heavy Docker containers on embedded hardware, and you absolutely cannot stream 24/7 raw in-car audio to a cloud server (a massive privacy liability). Qdrant Edge solves this. It acts like the SQLite of vector databases. You install it as a simple Python library: pip install qdrant-edge-py And open a directory on disk as a shard: from qdrant_edge import EdgeShard, EdgeConfig, EdgeVectorParams, Distance config = EdgeConfig( vectors={ "audio_embedding": EdgeVectorParams(size=1024, distance=Distance.Cosine) } ) shard = EdgeShard.create("./qdrant-edge-shard", config) You don’t require a daemon or Docker. The entire vector database lives inside the memory space of your Python program. It’s not like we don’t have a ton of other alternatives; we do, but the reason I chose Qdrant Edge is because of three main reasons: Absolute Privacy (vs. Cloud DBs): Vector matching happens entirely in-process. Audio is captured, embedded, matched, and discarded locally without ever leaving the vehicle. Built for the Edge (vs. Local Server DBs): Unlike running local instances of ChromaDB or Elasticsearch which drain RAM and CPU, Qdrant Edge relies on a highly optimized Rust core, making it perfect for low-power edge devices. True Database Features (vs. FAISS / NumPy): While you could hold arrays in memory with FAISS, you lose metadata. Qdrant Edge lets us attach metadata (like sound_type or severity), filter dynamically, and persist to disk effortlessly. . What We Are Actually Building The system has one job, and to do it as best as possible, it needs to listen to audio coming from a car’s microphone, and detect if a distress call sounds like a scream, a crash, shattering glass or an emergency siren — and send an alert to a Telegram contact. The whole pipeline looks like this: Capture live microphone audio in overlapping 1-second chunks. Run each chunk through YAMNet to produce a 1024-dimensional embedding. Search the Qdrant Edge shard for similar sounds using cosine similarity. Apply temporal smoothing: only trigger an alert if we get 3 hits within 5 seconds. Fire a Telegram message with the detected sound type, severity, and timestamp. The raw audio never leaves the device. All processing happens locally, in real time. Architecture Diagram Architecture The left side (provisioning) runs once. You download the ESC-50 sound dataset, embed every distress sound with YAMNet, and store those vectors in the Qdrant Edge shard. That shard then sits on disk indefinitely. The right side (runtime) runs continuously. Every 500 milliseconds of new audio produces a fresh embedding, searches the shard, and feeds into a sliding window detector. The shard file is the same one written during provisioning — there is no synchronization step, no replication, no cache warming. It just opens and reads. When you run python main.py , this is what start-up looks like: Spinning up the project The Audio Pipeline Capture The microphone is read using sounddevice. The capture runs in a background thread and puts overlapping chunks into a thread-safe queue. self._stream = sd.InputStream( samplerate=16000, channels=1, dtype="float32", blocksize=int(16000 * 0.01), # 10ms blocks callback=self._callback, ) The key design here is the sliding window. Each chunk is 1 second of audio, but we emit a new chunk every 500 milliseconds. That 50% overlap means a sudden sound that starts in the middle of a window still gets fully represented in the next one. You do not miss events that fall on boundaries. Preprocessing YAMNet has specific requirements: mono audio, 16kHz sample rate, float32 values between -1.0 and 1.0. The preprocessor enforces all of this regardless of what the microphone delivers. Why YAMNet and how does it fit ? YAMNet — Yet Another Mobilenet Network — is a pre-trained audio classification model from Google trained on the AudioSet dataset, which contains over 2 million human-labelled audio clips across 521 sound classes. The model is publicly available on TensorFlow Hub and weighs about 3 MB. It takes raw float32 waveform as input and produces three outputs: class probability scores, intermediate embeddings, and a log-mel spectrogram. We are not bothered about the class scores. We want the embeddings. import tensorflow_hub as hub model = hub.load("https://tfhub.dev/google/yamnet/1") _, embeddings, _ = model(waveform_tensor) Mean-pool over time frames to produce a single (1024,) vector embedding = tf.reduce_mean(embeddings, axis=0).numpy() L2-normalize so cosine similarity equals dot product norm = np.linalg.norm(embedding) if norm > 0: embedding = embedding / norm That 1024-dimensional vector captures the acoustic fingerprint of the sound. A scream and a car crash are going to produce very different fingerprints from someone humming or a two-way conversation. That difference is exactly what Qdrant Edge exploits during search. The reason to pick YAMNet over a general-purpose audio embedding approach is that it was trained specifically to understand audio events, not music or speech. It already knows what an emergency siren sounds like as a concept. We are just using its internal representation as a similarity signal. Building the Sound Library (Indexing) Before the detector can run, we need to populate the Qdrant Edge shard with reference embeddings. This is the one-time provisioning step. We use the ESC-50 dataset: 2000 environmental sound clips across 50 classes, all available for free. We pick the classes that matter for emergency detection, embed each one with YAMNet, and store the result with metadata. ALERT_CLASSES = { "screaming": {"sound_type": "scream", "severity": "high"}, "glass_breaking": {"sound_type": "glass_break", "severity": "high"}, "siren": {"sound_type": "siren", "severity": "high"}, "gunshot": {"sound_type": "collision", "severity": "high"}, "car_horn": {"sound_type": "car_horn", "severity": "medium"}, "crying_baby": {"sound_type": "crying", "severity": "medium"}, } Each sound gets stored as a point in the shard with a payload that includes its alert_class (either “alert” or “negative”), sound type, and severity. The alert_class field gets a keyword payload index: shard.update( UpdateOperation.create_field_index("alert_class", PayloadSchemaType.Keyword) ) This is important. At query time, we do not just want the nearest vectors globally we want the nearest vectors that are actually alert-worthy sounds. The keyword filter on alert_class lets Qdrant Edge restrict the search space to distress sounds only, which dramatically reduces false positives from background noise. search_filter = Filter(must=[ FieldCondition( key="alert_class", match=MatchValue(value="alert"), ) ] ) results = shard.query( QueryRequest( query=Query.Nearest(embedding.tolist(), using="audio_embedding"), filter=search_filter, limit=5, with_payload=True, ) ) The Detection Logic Raw similarity scores are noisy. A single loud sound that happens to be similar to a scream is not an emergency. An engine backfire can produce a high score momentarily. The system needs to tell the difference between a spike and a sustained pattern. The detector uses a sliding time window. A “hit” is recorded whenever the top similarity score exceeds the threshold (0.80 by default and different thresholds exist for different sounds and can be tweaked based on requirements). I have set value for the various sounds accurately here. An alert is only confirmed if at least 3 hits occur within a 5-second window. self._hit_window.append((time.time(), event)) now = time.time() while self._hit_window and (now - self._hit_window[0][0]) > self._window_secs: self._hit_window.popleft() recent_hits = len(self._hit_window) if recent_hits >= self._hits_required: self.total_alerts += 1 event.hit_count = recent_hits self._hit_window.clear() self._on_alert(event) This catches the real terminal output from a live test session. Three hits across 1.5 seconds of audio confirmed — and an alert is fired. The Telegram message arrives within a second of the third detection hit. Context-Aware Thresholding and Amplitude Gating Not all sounds behave the same way, so a global threshold doesn’t work in practice. The system uses per-class threshold overrides to balance sensitivity and precision: Amplitude Gating : Before a chunk is even embedded, we calculate its RMS volume. If it’s too quiet (e.g., standard AC hum or road noise), it gets dropped immediately. This saves compute cycles and prevents silent background noise from triggering false positives. Strict Mode for Broadband Sounds : Sounds like a car horn or a siren share acoustic similarities with heavy wind or engine revs. To prevent false alarms, these classes require a very strict similarity score (e.g., 0.90). Sensitive Mode for Impulsive Sounds : An emergency scream or a gunshot might be muffled or extremely brief. For these, we lower the similarity threshold to 0.80 and only require 2 hits instead of 3. This ensures the system reacts instantly to sudden, violent impulses while remaining stubbornly resistant to ambient noise. Flying Blind on the Edge: Adding SigNoz Observability While experimenting with the system on my desk, everything worked flawlessly. But then I realized a massive operational problem: what happens when this is deployed in a moving car? Edge deployment means you are effectively flying blind. If a device starts overheating and thermal throttling, or if a rattling car part starts triggering hundreds of “glass break” false positives, I wouldn’t know. I can’t exactly SSH into a vehicle driving down the highway to read the terminal output. I needed a way to monitor the system’s health, latency, and detection rates without ever recording or sending raw audio to the cloud (which would violate the entire privacy-first architecture). To solve this, I instrumented the pipeline using OpenTelemetry and routed the data to SigNoz . Because we are only sending metadata (timestamps, processing durations, and similarity scores), the privacy of the vehicle remains completely intact. Tracking the Bottlenecks: The 500ms Window The detector processes audio in 500ms chunks. If the hardware takes longer than 500ms to extract features, run the YAMNet embedding and search Qdrant, and the pipeline will buffer and eventually collapse. By wrapping my pipeline in OpenTelemetry spans, I could instantly see exactly where the computational time was going. Trace flow at SigNoz As you can see in the trace above, the total chunk processing time (sos.process_chunk) is clocking in around 24ms. The preprocessing (sos.preprocess) is near instantaneous (0.26ms). The heavy lifting is done by the YAMNet embedding (sos.embed), taking about 22.8ms. The Qdrant Edge similarity search (sos.vector_search) takes a mere 0.29ms. This gave me the confidence that the system operates well within the 500ms deadline, even on constrained hardware. Fleet-Wide Metrics Without Raw Data metrics at SigNoz Beyond just latency, I needed to track behavior . I created custom metrics in SigNoz to count sos.detection.hits (every time a sound breaks the similarity threshold) and sos.alerts.sent (when the temporal smoothing triggers a confirmed SOS). If I deployed this to a fleet of 5,000 cars and noticed that one specific vehicle model was suddenly generating massive spikes in sos.detection.hits for “car_horn” without triggering real alerts, I would instantly know there was an anomaly — perhaps a mechanical noise specific to that chassis. I could then adjust the SIMILARITY_THRESHOLD for that specific fleet via an OTA config update. Adding SigNoz took the project from being a “cool local script” to a resilient, production-ready architecture capable of being managed at scale. The Telegram Alert The alert module handles two channels: a Telegram HTTP call and a system audio beep. A cooldown timer (30 seconds by default) prevents spam in case of a sustained alarm. message = ("IN-CAR SOS ALERT\\n\\n" f"Severity: {event.severity.upper()}\\n" f"Sound Detected: {event.sound_type}\\n" f"Match Score: {event.score:.4f}\\n" f"Hits in Window: {event.hit_count}\\n" f"Time: {ts}\\n\\n" "Powered by Qdrant Edge · On-device detection" ) response = requests.post( f"https://api.telegram.org/bot{TELEGRAM_BOT_TOKEN}/sendMessage", json={"chat_id": TELEGRAM_CHAT_ID, "text": message}, timeout=10, ) One design decision worth calling out here: send the message as plain text, not Markdown. Sound type names like car_horn or glass_break contain underscores, and Telegram’s Markdown parser treats them as italic delimiters. Plain text avoids the whole problem. This is what the alerts actually look like arriving on your phone: Tech Stack Audio capture — sounddevice Low-latency microphone stream with callback-based processing Preprocessing — librosa + NumPy Audio resampling and normalization Embeddings — YAMNet (TF Hub) ~3 MB model trained on audio events with 1024-dimensional output Vector search — qdrant-edge-py In-process vector search with no server dependency and persistent on-disk storage Alerting — requests (Telegram Bot API) Simple and reliable notifications without requiring app installation Configuration — python-dotenv Secure credential management using .env files instead of hardcoded secrets The entire stack runs on an Apple M4 with Metal GPU acceleration for TensorFlow. On Linux edge hardware (a Raspberry Pi 5 or an Orin NX), it runs on CPU and still meets real-time requirements comfortably. Design Decisions Worth Thinking About Why filter by alert_class instead of just using a threshold? A high cosine similarity score means the query sound is close to something in the database. But “close to” includes close to negative examples, too. Engine noise can be acoustically similar to a siren at certain frequencies. By only searching within the alert class of points, we avoid ever scoring against the negative examples at all. The filter happens inside the ANN search itself, not as a post-processing step; Qdrant Edge handles this efficiently because of the keyword index. Why 3 hits in 5 seconds and not just 1? A single high-scoring hit almost certainly means a relevant sound was detected. But a transient event like a door slamming or a sharp noise from the road can spike above 0.80 for one chunk and then disappear. Three confirmed hits in 5 seconds means the sound is sustained, which is exactly what distinguishes a real emergency (a sustained siren, ongoing screaming, an alarm) from a false trigger. Why mean-pool YAMNet’s frame embeddings? YAMNet operates on roughly 0.48-second frames and produces one embedding per frame. A 1-second audio chunk produces about 2 embeddings. Mean-pooling collapses these into a single representative vector. The alternative storing multiple vectors per chunk and aggregating search scores — would complicate the indexing and search logic significantly. Mean-pooling is simpler and works well because the acoustic character of a 1-second distress sound is consistent across frames. Why alert-usage and not alerter? Authenticity. A system built by a human engineering team in a real product looks like real engineering decisions, not auto-generated code. File names, variable names, and comment style all contribute to whether the codebase feels like something someone has built or something that was generated. Numbers YAMNet model size: ~3 MB Qdrant Edge shard (360 indexed sounds): ~8 MB End-to-end latency per chunk: ~40ms (M4 with Metal) Alert delivery (Telegram): <1 second after confirmation Memory footprint (total process): ~320 MB Audio data sent off-device: 0 bytes The 360-vector shard covers 7 alert sound categories and 9 negative categories, with up to 40 clips per alert class and 20 per negative class. That is small enough to initialize in under 10 milliseconds and fits entirely in the L2 cache on modern ARM processors. What’s Next? The biggest gap right now is the alert routing. Telegram works well for a demo, but production-grade alerting would want to integrate with emergency service APIs, push to a fleet management dashboard, or trigger an automated call. The detector’s on_alert callback interface is already designed for this — you swap the Telegram handler for anything else. Data Augmentation: Simulating the In-Car Environment While ESC-50 provides a great baseline, real-world emergency sounds don’t happen in a soundproof studio — they happen over the rumble of an engine. To bridge this gap, we implemented custom dataset ingestion. We took raw, real-world files of screams and gunshots and programmatically augmented them during the indexing phase. Before passing these custom samples into YAMNet, the preprocessor dynamically overlays the audio on top of a base track of a car engine standing idle. It then indexes both the clean version and the engine-augmented version into the Qdrant Edge shard. This means our vector database is explicitly populated with the acoustic fingerprints of emergencies happening inside a running vehicle , dramatically boosting real-world recall without needing to collect thousands of hours of ƒdashcam audio. Finally: speaker identification. Right now the system detects distress sounds generically. Adding a voiceprint comparison layer using the car owner’s enrolled voiceprint as a reference embedding stored in the same Qdrant Edge shard would allow the system to distinguish a passenger screaming at a horror movie podcast from the driver screaming in a genuine emergency. The Core Idea Every technology decision in this system points in the same direction. Audio stays on the device. The model runs locally. The vector database runs in-process. The only outbound call is the alert itself, and that is the entire point. Qdrant Edge made all of this possible without any infrastructure gymnastics. The same tool that powers production-grade semantic search at scale also runs in 8 megabytes of disk space inside a Python process on embedded hardware. That is the part that is easy to miss when you first see it. This is not a simplified version of vector search. It is the full thing, just without the server. References Qdrant Edge Documentation YAMNet on TensorFlow Hub ESC-50 Dataset (Karol Piczak) Telegram Bot API sounddevice documentation AudioSet — Google Research How I Built a Real-Time In-Car SOS Detection System With Qdrant Edge, SigNoz, and YAMNet was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
- ServiceNow CMO to Join OpenAI
ServiceNow CMO to Join OpenAI The Information
- Google Introduces Gemini 3.5 Flash Low to Help Antigravity Users Maximise Usage
Ever since Google restructured its Gemini plans to a compute-based usage system, it has struggled to provide users of its agentic coding platform, Antigravity, with enough tokens. The Mountain View-based tech giant, last week, announced a shift from the message-based usage to one that measures the number of tokens exhausted. Once the new system was rolled out, users b...
- Faire doubles PR throughput with Cursor Cloud Agents
Faire achieves doubled PR throughput with Cursor Cloud Agents.
- Building the AI-ready data foundation – join the webinar
BITanium, an IBM business partner, will show how companies simplify hybrid data integration, improve AI readiness and accelerate trusted AI outcomes with IBM watsonx.data.
Score: 25🌐 MovesMay 26, 2026https://www.itweb.co.za/article/building-the-ai-ready-data-foundation-join-the-webinar/P3gQ2MGAZAjvnRD1 - US-DATA Expands AI Data Annotation Services
US-DATA Expands AI Data Annotation Services USA Today
Score: 25🌐 MovesMay 26, 2026https://www.usatoday.com/press-release/story/33076/us-data-expands-ai-data-annotation-services/ - Dow Jones Leader Amazon, AI Stock Quanta, Monster Beverage In Or Near Buy Zones
Dow Jones leader Amazon stock and AI stock Quanta Services are in or near buy zones in the ongoing stock market rally. The post Dow Jones Leader Amazon, AI Stock Quanta, Monster Beverage In Or Near Buy Zones appeared first on Investor's Business Daily .
Score: 24🌐 MovesMay 26, 2026https://www.investors.com/stock-lists/stocks-near-a-buy-zone/dow-jones-amazon-stock-amzn-stock-ai-stock/ - Boston cybersecurity startup 7AI is going on a hiring binge
Boston cybersecurity startup 7AI is going on a hiring binge The Boston Globe
Score: 24🌐 MovesMay 26, 2026https://www.bostonglobe.com/2026/05/26/business/7ai-liv-dior-headcount-hiring/ - Trump Shares AI Picture of Him Hunting RINOs
Trump Shares AI Picture of Him Hunting RINOs Newsweek
Score: 24🌐 MovesMay 26, 2026https://www.newsweek.com/trump-shares-ai-picture-of-him-hunting-rinos-11992729 - Echelon Philippines 2025 – The future is Filipino: Opportunities in AI
At Echelon Philippines 2025, Carlo Almendral, CEO and Co-Founder of AIFirst, delivered a compelling keynote speech on the second day of the event, painting an optimistic picture of the Philippines’ place in the rapidly evolving AI landscape. Drawing on the country’s unique strengths, Almendral highlighted the immense potential the Philippines holds in embracing and advancing […] The post Echelon Philippines 2025 – The future is Filipino: Opportunities in AI appeared first on e27 .
Score: 24🌐 MovesMay 26, 2026https://e27.co/echelon-philippines-2025-the-future-is-filipino-opportunities-in-ai-20260526/ - These AI bots want to help fans navigate World Cup host cities
When soccer fans head to the FIFA World Cup starting in June, they’ll have a new option for finding things to do, places to stay, and ways to get around in various host cities: artificial intelligence . Some visitors will likely turn to general-purpose AI tools like ChatGPT and Claude, with recent studies finding they’ve become popular travel-planning tools despite the risk they’ll present outdated or false information. But multiple host cities are also deploying specially developed AI assistants and virtual concierges they say will deliver curated, relevant, and up-to-date facts and guidance for visitors. Frisco, Texas—home to Toyota Stadium, which is serving as a World Cup base camp for Sweden’s national team—has worked with the AI travel planning service GuideGeek on an AI assistant called Frankie that can answer questions about hotels, restaurants, shopping, and activities in the Dallas-area city. The bot launched last year after a couple of months of work, ensuring it could give correct and comprehensive answers based on data sources like the official Visit Frisco tourism website, says Cori Powers, director of marketing and communications for Visit Frisco. “We really wanted to ensure that it was conversational and fun and would make trip planning to Frisco convenient,” says Powers. Recently, Powers says, the bot has seen a rise of questions about World Cup planning, along with other questions related to summer vacations. Noting the questions users ask Frankie has helped the organization add relevant copy to its website—which in turn feeds back into Frankie—and social media channels. “One of the biggest values for the tourism boards is identifying where those content gaps are,” says Greg Oates, director of AI advocacy at GuideGeek and its parent organization, Matador Network . “If a tourism board has seen that a lot of people are asking about something specific and it’s not being answered in the website, then they can update that content or expand on that.” For better or worse, GuideGeek’s city-specific bots are designed to steer off-topic questions back to their sponsoring cities, which means Frankie reflects even some questions about the greater Dallas area back to answers about Frisco itself. But the bots also have features that ordinary tourism websites don’t, including the conversational interface, map integrations to highlight relevant sites, and the ability to serve up relevant images. Additionally, GuideGeek bots, which serve more than 50 locations and brands around the world from Aruba to Manitoba, can answer questions in dozens of languages. “If you’re coming from somewhere and English isn’t your native language, you just talk to GuideGeek in whatever your language might be,” Oates says. “GuideGeek understands that [and] will respond in kind.” Already, Visit Frisco has seen a burst of queries in languages like Spanish, German, and Mandarin. And GuideGeek’s multilingual capability has also proven useful in New York City, where NYC Tourism + Conventions has deployed two GuideGeek-powered bots: Ellis , targeting business event planners, and Libby , aimed at tourists and travelers visiting the city. The tourism organization deployed Libby last year, motivated in part by the World Cup, and the fact that while its website is only available in five languages, GuideGeek’s AI can support more than 60. Libby, which is available through the Tourism + Conventions website and through WhatsApp , quickly proved popular, says Nancy Mammana, chief marketing officer at NYC Tourism + Conventions. “When we launched it in June, it really started to catch fire quickly, and it’s become a very important channel for us,” she says. “We’ve seen over 45,000 conversations happen with the tool in 68 languages from 178 countries, and over 122,000 queries, which is great.” Libby, which is advertised with QR code-embedded marketing material throughout the city, has even been embraced by locals for some events, like Restaurant Week, Mammana says. It also won’t be the only AI bot available to help navigate the New York area around the World Cup, which in addition to potentially heavy crowds will see changes to normal transit patterns , along with special deals at restaurants and exhibits at area museums. An “Official NYNJ World Cup Concierge” will also be available with the backing of the official FIFA World Cup New York New Jersey host committee , built with a company called Neurun that got its start building AI guides for running events like marathons. The AI concierge, which will be accessible through the host committee website and other websites that embed its web page widget with host committee approval, is designed to be a single “official source of truth” for the World Cup events, says Bruce Revman, cohost city manager of the FIFA World Cup 26 New York New Jersey host committee. That means that it will have access to up-to-date transit info, highlighted through an integration with Google Maps, along with other verified information about what’s going on in the area during the World Cup events. Users will also be able to ask for general New York City information, like opening hours at area attractions, or use the tool to locate places to watch World Cup games and find special deals available during the tournament, Revman says. In addition to testing its AI concierges by hand, Neurun deploys additional AI agents that pose questions of the bots and grade and record their answers, says Neurun’s cofounder and CEO, Cade Netscher. “It’ll record the activity that it does, so we can watch it, ask different questions, see what happens, make sure it looks appropriate,” he says. “And then we can fill in the gaps.” Like Libby, the concierge is likely to have uses beyond the World Cup. Revman says it’s expected to be promoted around events like Sail 4th 250 , a celebration of the country’s 250th anniversary this July focusing on tall ships. And while AI travel planning tools are sometimes criticized for taking the human element out of vacationing, replacing personal research and expert advice with computer-generated itineraries and fact sheets, Revman emphasizes that questions will be based on official information derived from human expertise, whether users are asking about security protocols, sightseeing options, or travel logistics. “It’s been a fun time, working with the host committee and their partners in this,” says Netscher. “You see AI headlines—everyone’s terrified of AI replacing human connection and everything—and we think with this technology we can leverage AI to enhance human connection.”
- Underground area in central Seoul to be repurposed into space showcasing AI, hologram technologies
A rendering of how the underground space will be repurposed by Creative MUT [CREATIVE MUT] The content solutions company Creative MUT will repurpose an unused underground area in central Seoul into a space that showcases Korean content through AI and hologram technologies. The Seoul Metropolitan Government selected Creative MUT as the private operator to develop the unused space underneath City Hall Station and Euljiro 1-ga Station, the company announced on Tuesday. Related Article BTS concert film exhibition opens in Seoul this fall Immersive 'Übermensch' exhibition lets G-Dragon fans step into artist's creative world KCON Japan 2026 to showcase K-beauty, K-food, and K-content innovations The company aims to finish development and open the area in October. The unused space is a 3,261-square-meter (35,101-square-foot) plot of land with a floor-to-ceiling height of 3.5 to 4 meters (11.5 to 13 feet). Spanning City Hall Station and Euljiro 1-ga Station, the area has been inaccessible to the public for the past 40 years. Creative MUT will be in charge of content planning, as well as investing in and operating the repurposed site. The company plans to use its capabilities in hologram technology and AI-based interactive content to host exhibitions and programs. Meanwhile, Seoul will be responsible for safety-related construction, including ventilation, firefighting systems and escape systems, and the electrical and communication infrastructure. An example of the exhibitions that will be held at the repurposed underground space of City Hall Station and Euljiro 1-ga Station [CREATIVE MUT] The repurposed area will be divided into three zones: one dedicated to AI-generated media art and interactive exhibitions; another hosting hologram-based virtual performances and photo zones inspired by K-pop and K-dramas; and a third where visitors can experience content through extended reality (XR) and virtual reality. The space will also serve as a testbed for local startups that wish to trial their AI, hologram or XR technologies. Creative MUT expects around 1 million people to visit the venue each year, based on estimates that around 90,000 people visit the City Hall Station and Euljiro 1-ga Station every day. “The space where the noise of the city disappears, located 335 meters underground in central Seoul, will showcase Korea's AI and hologram technologies and become a place visited by K-culture fans around the world,” said Kim Tae-hwan, the CEO of Creative MUT. “We will show citizens of Seoul and those around the world how technology meets culture by using our knowledge in entertainment and technology, which we have built in global markets.” BY LEE TAE-HEE [lee.taehee2@joongang.co.kr]
- Jaguars QB Trevor Lawrence has no plans to cut his long locks despite team's deep fake video
Jaguars QB Trevor Lawrence has no plans to cut his long locks despite team's deep fake video Houston Chronicle
Score: 23🌐 MovesMay 26, 2026https://www.houstonchronicle.com/sports/article/jaguars-qb-trevor-lawrence-has-no-plans-to-cut-22277646.php - AI Chatbot Pricing Comparison: Here's What You Get When You Pay
Are AI subscriptions worth it? It depends.
Score: 22🌐 MovesMay 26, 2026https://www.cnet.com/tech/services-and-software/upgrading-your-ai-chatbot-heres-how-much-itll-cost-you/ - Inside OpenAI’s New Founder Experience Team
A look at OpenAI's new Founder Experience team led by Laura Modiano. The team plans to work alongside startups to offer tailor-made support to founders.
Score: 22🌐 MovesMay 26, 2026https://www.forbes.com/sites/timkeary/2026/05/26/inside-openais-new-founder-experience-team/ - Got a Backlog of Long-Form Articles? Spotify Can Now Read Them Aloud
Got a Backlog of Long-Form Articles? Spotify Can Now Read Them Aloud PCMag
Score: 22🌐 MovesMay 26, 2026https://www.pcmag.com/news/got-a-backlog-of-long-form-articles-spotify-can-now-read-them-aloud - Microsoft cheat sheets: Dive into Windows, Office, and Copilot
Need to get up to speed on the latest features in Excel? Wrestling with an old version of Word? Looking to get more out of Windows 11? Computerworld ’s cheat sheets are easy-to-use guides to help you navigate Microsoft’s core productivity software. Here’s a one-stop resource where you can find in-depth stories on several generations of Word, Excel, PowerPoint, and Outlook for Windows, focusing on what’s new in each major release. We’ve also got guides for Windows itself, as well as Microsoft Teams, OneDrive (both in Windows and on the web), OneNote, Loop, Whiteboard, Forms, Visio, Planner, and Power Automate. Microsoft’s subscription-based office suite, called Microsoft 365 , is continually updated with new features, so we periodically refresh the cheat sheets for the “365” versions of Word, Excel, PowerPoint, Outlook, and other apps in the suite. But some companies and individuals will likely stay on older versions of the non-subscription software (Office 2021, for example) for some time to come, so we’ve got cheat sheets for several generations of those products as well. The biggest change in both Microsoft 365 and Windows in recent years is the widespread integration of Microsoft’s generative AI assistant, Copilot. We’ve got tips to help you get the most out of that tool too, with more on the way. Windows, Office, and Copilot tutorials and tips Microsoft Copilot Windows 10 and 11 Microsoft 365 apps Office 2021 and 2024 Office 2016 and 2019 Office 2013 Office 2010 Windows 8 Microsoft Copilot Microsoft Copilot tips: 9 ways to use Copilot right The free version of Microsoft’s generative AI chatbot is available in a standalone app, in the Edge browser, and on the web. Here’s how to make the most of it. New: 11 cool things Copilot can do in Excel As Microsoft ramps up Copilot’s capabilities in Excel, the AI tool is becoming genuinely useful for spreadsheet work. New: Copilot Chat: Your hub for document creation and analysis Anyone with a Microsoft 365 account can use new AI agents in Copilot Chat to jump-start Word, Excel, and PowerPoint documents — and some users can enlist a special agent to analyze their M365 files. Here’s how. Microsoft Copilot can boost your writing in Word, Outlook, and OneNote — here’s how Copilot integration in Microsoft 365 apps makes it a snap to generate first drafts, revise text, and get instant summaries for long docs or email threads. Here’s how to use Copilot for writing assistance in Word, Outlook, and OneNote. How to curb hallucinations in Copilot (and other genAI tools) Generative AI chatbots like Microsoft Copilot make stuff up all the time. Here’s how to rein in those lying tendencies and make better use of the tools. Windows 10 and 11 Windows 11 cheat sheet To a great extent, Windows 11 looks and works like Windows 10, but there are several minor differences that take some getting used to. We cover all the important changes here, including Copilot integration and new three- and four-finger touchscreen gestures. Windows 10 cheat sheet Windows 10 is no longer receiving updates , but if you’re still using this reliable workhorse, here’s a guide to the key features. Don’t miss our list of handy gestures and shortcuts for Windows 10. Microsoft OneDrive cheat sheet: Using OneDrive in Windows If you have Windows 10 or 11, you have OneDrive. Here’s how to back up, sync and share files in OneDrive and OneDrive for Business on the Windows desktop. More tips for Windows 10 and 11 How to curb hallucinations in Copilot (and other genAI tools) Microsoft Copilot tips: 9 ways to use Copilot right Updated: 8 ways to be more productive in Windows 11 15 ways to speed up Windows 11 Updated: How to protect your privacy in Windows 11 How to repair Windows 10 or 11 in 4 steps How to handle Windows 10 and 11 updates How to protect Windows 10 and 11 PCs from ransomware How to share a Windows 10 or 11 PC 12 classic but essential (and free!) utilities for Windows 10 and 11 18 ways to speed up Windows 10 How to protect your privacy in Windows 10 30+ free and cheap apps for Windows 10 Microsoft 365/Office 365 apps Word for Microsoft 365 cheat sheet Learn to use the best features introduced in Word for Microsoft 365 in Windows over the past several years. This story covers features introduced in Word 2016, 2019, 2021, and 2024, plus several more exclusive to Microsoft 365 subscribers — and to those with a Microsoft 365 Copilot license. Related: Handy Word keyboard shortcuts for Windows and Mac Microsoft Copilot can boost your writing in Word, Outlook, and OneNote — here’s how Updated: 10 quick productivity tips for Microsoft 365 mobile apps Microsoft Word for Android (and iOS) cheat sheet Excel for Microsoft 365 cheat sheet Learn about the most important features introduced in Excel over the past several years, with an emphasis on those exclusive to Microsoft 365 subscribers — and to users with a Microsoft 365 Copilot license. Related: Handy Excel keyboard shortcuts for Windows and Mac New: 11 cool things Copilot can do in Excel Updated: 10 quick productivity tips for Microsoft 365 mobile apps How to use PivotTables and PivotCharts in Excel How to use slicers in Excel Excel basics: Get started with tables Excel basics: Get started with charts and sparklines How to use Excel formulas and functions How (and why) to use conditional formatting in Excel How to use Excel macros to save time and automate your work 10 spiffy new ways to show data with Excel Excel: Your entry into the world of data analytics How to use Excel as a data visualization tool 8 simple ways to clean data with Excel PowerPoint for Microsoft 365 cheat sheet Learn to use the best features introduced in PowerPoint for Microsoft 365 in Windows over the past several years. This story covers the major features introduced in PowerPoint 2016, 2019, 2021, and 2024, plus several more exclusive to Microsoft 365 subscribers — and to those with a Microsoft 365 Copilot license. Related: Handy PowerPoint keyboard shortcuts for Windows and Mac Updated: 10 quick productivity tips for Microsoft 365 mobile apps Outlook for Microsoft 365 cheat sheet Discover all the major features introduced in Outlook 2016, 2019, 2021, and 2024, plus more exclusively for Microsoft 365 subscribers — including a simplified Ribbon that shows only the most commonly used commands. Related: Handy Outlook keyboard shortcuts for Windows and Mac Microsoft Copilot can boost your writing in Word, Outlook, and OneNote — here’s how Updated: 10 quick productivity tips for Microsoft 365 mobile apps Use Outlook’s new calendar board view to organize your work 8 Outlook add-ins to enhance collaboration How to filter Outlook emails on all your devices Updated: How to check your co-workers’ schedules in Outlook and Teams How to work across time zones in Outlook Microsoft Loop cheat sheet Microsoft’s new Loop app provides shared workspaces where teams can collaborate. Our cheat sheet shows you how to use the Loop app. How to use Loop components in Microsoft 365 apps What makes Loop particularly useful is the ability to collaborate on content snippets called Loop components across multiple Microsoft 365 apps. Here’s how to use Loop components in Outlook, Teams, and other M365 apps. Updated: Microsoft Teams cheat sheet: How to get started Microsoft’s answer to Slack and Zoom, Teams provides group messaging, voice and video calls, and useful integrations with other Microsoft 365 apps. Here’s how to get set up in Teams and find your way around. Related: 28 power user tips for Microsoft Teams 14 best practices for Microsoft Teams video meetings The 10 best new Microsoft Teams meeting features Updated: How to check your co-workers’ schedules in Outlook and Teams How to have Teams meetings with people outside your organization 18 Microsoft Teams apps for content collaboration and management Microsoft OneNote cheat sheet Part of Microsoft’s Office suite and built into Windows 10 and 11, OneNote is a robust note-taking app that is also available as a free standalone product. Here’s how to get up and running with OneNote. Related: Microsoft Copilot can boost your writing in Word, Outlook, and OneNote — here’s how Updated: 10 quick productivity tips for Microsoft 365 mobile apps Microsoft OneDrive cheat sheet: Using OneDrive for Web OneDrive for Web lets you save, access, share, and manage your files in the cloud using your favorite browser. Learn how to use the web interface — and Copilot AI with it — for a big productivity boost. Updated: Microsoft Forms cheat sheet: How to get started Online forms help you conduct research, collect feedback, test knowledge, and more. Here’s how to use Microsoft Forms to create surveys, feedback forms, quizzes, and other interactive forms. Microsoft Visio cheat sheet: How to get started Visio in Microsoft 365 is an excellent tool for creating custom diagrams to illustrate concepts that are difficult to explain through text. Here’s how to use it. 13 tips to get the most out of Microsoft Whiteboard For Microsoft 365 users, it’s worth adding Microsoft Whiteboard to your collaboration playbook. Here’s how your team can make the most of this digital whiteboard tool. Updated: Microsoft Planner cheat sheet Planner gives Microsoft 365 users a built-in task-management tool that small teams can use to track plans, tasks, and progress. Here’s our guide to using Planner on the web and within Microsoft Teams. Microsoft Power Automate: How to get started With Power Automate, you can create automated workflows for a wide range of business tasks across multiple apps and services — no coding required. Here’s how to get up and running, along with tips for creating reliable automations. SharePoint Online cheat sheet Learn how to find your way around SharePoint Online (the Office 365 version of SharePoint), create sites, share and manage documents, work with calendars, integrate with Outlook and more. Then go beyond the basics in 5 tips for working with SharePoint Online . More tips for Microsoft 365/Office 10 highly useful add-ins for Microsoft Office 5 collaboration tools that enhance Microsoft Office Updated: 5 steps to repair Microsoft Office Office 2021 and 2024 Office 2021 and 2024 cheat sheet Microsoft 365 may get all the attention, but the classic Microsoft Office suite also gets useful additions in every release. Here’s how to use the best new features in Office 2021 and Office 2024. Office 2016 and 2019 Word 2016 and 2019 cheat sheet Learn how to use Word’s live collaborative editing features, Tell Me and Smart Lookup, and the new Translator pane in Word 2019. Also included is a list of handy keyboard shortcuts for Word 2016 and 2019 . If you just want to know where to find various commands on the Ribbon, download our Word 2016 and 2019 Ribbon quick reference . Excel 2016 and 2019 cheat sheet Now updated for Excel 2019, our guide covers several useful chart types introduced in Excel 2016 and Excel 2019 for Windows, as well as how to use several impressive new data analysis tools. We’ve also got a list of handy keyboard shortcuts in Excel , as well as the Excel 2016 and 2019 Ribbon quick reference . PowerPoint 2016 and 2019 cheat sheet Like Word and Excel, PowerPoint 2016 and PowerPoint 2019 for Windows offer Tell Me, Smart Lookup, live collaborative editing and a slew of new chart types. We cover all that plus some handy features introduced in PowerPoint 2019 — not to mention our list of keyboard shortcuts for PowerPoint and the PowerPoint 2016 and 2019 Ribbon quick reference . Outlook 2016 and 2019 cheat sheet Outlook 2016 for Windows has been enhanced with Smart Lookup, Tell Me, and features to help you find files you want to attach and keep a tidy inbox. And don’t miss our list of keyboard shortcuts for Outlook 2016 and 2019 and the Outlook 2016 and 2019 Ribbon quick reference . Office 2013 Word 2013 cheat sheet Among the major features introduced in Word 2013 are a Start screen, a Design tab, Read Mode, and OneDrive sync. Our guide covers how to use them all and provides handy keyboard shortcuts for Word 2013 . There’s also a Word 2013 Ribbon quick reference . SharePoint 2013 cheat sheet Learn the basics of navigating and using a SharePoint site, where to go to find some of the customization options, and 5 advanced SharePoint 2013 tips . Office 2010 Word 2010 cheat sheet Learn how to use Word 2010’s Navigation pane, image editing tools, text effects and other new features. Also see the list of handy keyboard shortcuts for Word 2010 and our Word 2010 Ribbon quick reference charts . Excel 2010 cheat sheet Excel 2010 introduces Sparklines, Slicers, and other enhancements to PivotTables and PivotCharts. Find out how to use those, along with keyboard shortcuts for Excel 2010 and our quick reference for finding your favorite commands on the Excel 2010 Ribbon. PowerPoint 2010 cheat sheet Learn how to use PowerPoint 2010’s multimedia editing tools, sharing options and other handy features. As usual, we’ve got keyboard shortcuts for PowerPoint 2010 and a guide to finding old PowerPoint 2003 commands on the PowerPoint 2010 Ribbon . Outlook 2010 cheat sheet The Ribbon was only half-present in Outlook 2007, but in Outlook 2010 it’s ubiquitous. Other notable changes include Conversation View to group email messages, Schedule View for scheduling meetings, and an enhanced search function. We show you how to use them all, provide some handy keyboard shortcuts for Outlook 2010 and detail where old Outlook 2003 commands are located in Outlook 2010 . SharePoint 2010 cheat sheet Unlike earlier versions of SharePoint, SharePoint 2010 is based on the Ribbon interface. Here’s how to find your way around and get started with a SharePoint site. Windows 8 Windows 8 cheat sheet Not many people are still using this nightmare of an operating system , which radically overhauled the classic Windows interface in an attempt to make it more like a mobile OS. Just in case, here’s help finding your way around. (But seriously, it’s way past time to upgrade to a newer OS.)
Score: 22🌐 MovesMay 26, 2026https://www.computerworld.com/article/1682358/microsoft-cheat-sheets-dive-into-windows-and-office-apps.html - How can this smart robot lawn mower learn your yard on its own?
How can this smart robot lawn mower learn your yard on its own? USA Today
Score: 22🌐 MovesMay 26, 2026https://www.usatoday.com/story/shopping/deals/home/outdoor/2026/05/25/dreame-robot-lawn-mower/90249958007/ - Ask Anna: He’s using AI to text me
Ask Anna: He’s using AI to text me Chicago Tribune
Score: 21🌐 MovesMay 26, 2026https://www.chicagotribune.com/2026/05/26/ask-anna-my-boyfriend-used-chatgpt-to-write-romantic-texts-2/ - Claude Code Free with Local AI
5 STEPS TO ZERO-COST CLAUDE CODE → Step 1 · Install Claude Code → Step 2 · Install Ollama → Step 3 · Pull the Right Model → Step 4 · Connect Claude Code to Your Local Model → Step 5 · Expand the Context Window to 64K Tokens Source: Image by the author. Claude’s standard subscription costs $17/month, while the Max plan hits $100/month. But even with an active subscription, context limits run out fast when you are working inside large codebases. The moment your session hits its ceiling, you’re stuck either waiting or paying more. The better way is that you can run Claude Code using a completely free, open-source model “ Qwen 3.5 9B” ,running locally on your own machine via Ollama. This means no cloud bills, no rate limits eating into your workflow and completely private code handling. Note: You still need to be on at least a Claude Pro plan to access and initialize the Claude Code tool itself; what this local setup eliminates is the heavy API model usage/inference costs. System Prerequisites Before diving into the terminal commands please ensure your hardware environment meets the baseline operational demands. Local LLMs are heavily dependent on system memory and compute capabilities. OS: Windows 10 or 11 (PowerShell 7+ recommended). RAM: 16GB minimum. If you have 8GB, the model will offload aggressively to system storage, causing massive performance lag. Storage: At least 10GB of free SSD storage space to safely host the weights and system dependencies. Step 1: Install Claude Code & Configure Your Path Open PowerShell and run the installation command for Claude Code. During the installation process, pay close attention to the output terminal. It will display a specific folder path where the executable is being stored, looking something like this: C:\Users\ \.local\bin Crucial Missing Step - The installer does not always automatically append this to your system environment variables. If you skip this, your terminal will throw a “command not found” or “path not recognized” error the first time you try to launch it. Press the Windows Key , type “Environment Variables” , and select Edit the system environment variables . Click the Environment Variables button at the bottom. Under User variables , select Path and click Edit . Click New and paste your path: C:\Users\ \.local\bin (make sure to replace with your actual Windows account name). Click Move Up to push it to the top of the list so it executes cleanly. Click OK to save and close out of the windows. Step 2: Install and Verify Ollama Download and run the Ollama installer for Windows. Ollama installer for Windows : https://ollama.com/download/windows Once the installation finishes, open a fresh PowerShell terminal and verify that the Ollama background service is responsive by typing: ollama If it successfully returns a list of available commands, it’s alive and ready. You can check if you have any models already downloaded by running: ollama list Source: Image by the author. Step 3: Pull and Test the Base Model For a standard 16GB RAM machine, Qwen 3.5 9B is the absolute sweet spot. It is highly capable at handling complex coding syntax, refactoring, and logical reasoning without choking your physical hardware. To download the model, run: ollama pull qwen3.5:9b (The download is roughly 6.6 gb, so give it a few minutes depending on your internet speed). Once finished, give it a quick sanity check to make sure it’s interacting properly : ollama run qwen3.5:9b Type a quick prompt (e.g., "Write a fast sorting function in Python"). Once it responds correctly, exit the model prompt using Ctrl + D or by typing /exit. Step 4: Creating a Custom Modelfile Here is a critical catch that almost every guide on the internet skips: If you just run the base model, Ollama caps its default context length at only 16,384 tokens . While 16K is fine for a basic chat interface, it is a death sentence for a terminal agent like Claude Code. Claude Code needs to read entire directory structures, parse multiple files simultaneously, and look through historical context blocks. If you use the default limit, the model will quickly lose track mid-task and start hallucinating or failing. Qwen 3.5 natively supports up to 256K tokens, so we are going to expand our local context to a comfortable 64K tokens . Open up a plain text editor (like Notepad) and paste the following two lines: FROM qwen3.5:9b PARAMETER num_ctx 65536 2. Save this file exactly as Modelfile (make sure it has no .txt file extension) inside your active working directory. 3. Back in your terminal, build your new, superpower-backed model by running: ollama create qwen3.5-9b-64k -f Modelfile Verify it compiled correctly by typing ollama list. Your new qwen3.5-9b-64k model should show up right at the top of the list. Note: If you are tight on storage space, you can safely remove the original base model using ollama rm qwen3.5:9b since your new custom build handles the heavy lifting. Step 5: Launch Claude Code with Your Local Model Now it’s time to link them up. Navigate to the local coding project you want to work on using the cd command, and launch Claude Code pointed directly to your high-context local model: ollama launch claude-model qwen3.5-9b-64k Claude Code will initialize right inside your project directory. It will prompt a quick security safety check asking if you trust the project folder — confirm it, and you’re officially in. How to Verify It’s Working: To ensure Anthropic’s cloud isn’t secretly receiving your requests and billing your account, type the following command inside the Claude Code interface: /models You will see a matrix showing that all model slots — including Sonnet, Opus, and Haiku — are actively pointing back to your local qwen3.5-9b-64k model. Your local machine is now driving the entire developer environment. Real-World Performance If you want to track how your hardware is coping while you code, open up a separate, parallel terminal window and run: ollama ps This utility gives you a live look at your active local models and their exact VRAM/RAM footprints. On a mid-tier 16GB system, expect your resources to average roughly 45% CPU utilization and 55% GPU utilization during intensive generation cycles. Because it’s running locally, the model will take longer to “think” when executing complex repository sweeps. For instance, if you paste a massive GitHub URL repository and ask it to clone the codebase, map its entire architectural layout, find critical code flaws, and calculate deployment specifications, it can take up to 10 or 11 minutes of continuous processing. But the trade-off is unmatched: the output is incredibly thorough, entirely private, completely unlimited, and costs exactly $0 in API fees . Troubleshooting Common Windows Challenges Source: Image by the author. Even if it appears that you have every aspect set up as intended, running local AI tools on Windows can at times put a wrench in your workflow. If your terminal offers you trouble, then here are the most typical technical hurdles and specifically how to fix them: 1. “Command Not Found” (Even After Updating the Path) If you have meticulously added C:\Users\ \.local\bin to your environment variables but typing claude still throws an error, don't panic. Windows terminals do not automatically refresh their system environment map in active sessions. Fix: Simply close your current PowerShell window entirely and open a brand-new instance. If you are using an integrated terminal inside an IDE like VS Code, you will need to restart the entire IDE for the changes to take effect. 2. The Invisible .txt Extension Trap When creating your Modelfile, Windows File Explorer loves to be "helpful" by secretly saving it as Modelfile.txt while hiding known extensions from your view. If Ollama tells you it can't find or read the file during the compile step, this hidden extension is almost always the culprit. Fix: Open File Explorer, click on View (or Options depending on your Windows version), and check the box for File name extensions . If you see Modelfile.txt, right-click it, select rename, and delete the .txt entirely so it stands alone as just Modelfile. 3. Sluggish Performance and Intense System Lag If your entire machine locks up, freezes, or responds at a snail’s pace while Claude Code is processing a request, your system memory is likely hitting a hard bottleneck. This happens when your hardware is forced to offload model weights from your fast VRAM/RAM onto your slower local system drive. Fix: Close heavy background apps like Google Chrome tabs, Discord, or Docker containers before launching your local model. If you are on a strict 16GB RAM limit and resource contention persists, consider dropping the context limit in your Modelfile from 65536 down to 32768 (32K). This still gives you double the default capacity while significantly reducing the active memory footprint. 4. The “Connection Refused” Ollama Error If Claude Code throws an error saying it cannot establish a connection to your local backend, the Ollama background host daemon has either crashed or hasn’t initialized yet. Fix: Look at your Windows taskbar system tray (the small arrow icon in the bottom right corner) and verify that the Ollama llama icon is visible. If it isn’t there, search for Ollama in your Windows Start Menu and launch it manually to wake up the local server API before trying to run Claude Code again. Thanks for sticking around to the end. If you found these checkpoints helpful, a few claps go a long way. Let me know your thoughts regarding this article in the comments below. I am a tech writer and the founder of AnalystHQ, where I simplify data analytics and AI workflows. If you want to dive deeper into practical data skills, you can check out my complete playbook, the SQL & Excel Data Analyst Course . Claude Code Free with Local AI was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.
Score: 20🌐 MovesMay 26, 2026https://pub.towardsai.net/claude-code-free-with-local-ai-ffbc92208006?source=rss----98111c9905da---4 - xAI told staff to stop mingling with Cursor employees, weeks after they started working together
Elon Musk’s xAI has told employees to limit their contact with staff from Cursor , the AI coding startup that SpaceX has an option to acquire for $60 billion. The directive came from James Burnham, xAI’s general counsel and former chief lawyer at the Department of Government Efficiency, according to Bloomberg. Burnham sent guidelines to […] This story continues at The Next Web
- Y Combinator founder Paul Graham says AI-written founder emails feel like being lied to
Y Combinator founder Paul Graham ignores emails clearly written by AI—they feel "like being lied to," he says. That's coming from one of OpenAI's earliest investors. Studies suggest his reaction is anything but unusual. The article Y Combinator founder Paul Graham says AI-written founder emails feel like being lied to appeared first on The Decoder .
Score: 20🌐 MovesMay 26, 2026https://the-decoder.com/y-combinator-founder-paul-graham-says-ai-written-founder-emails-feel-like-being-lied-to/ - AI when you buy
Shoppers browse the aisles of a supermarket in downtown Seoul on May 26, as the government announced it would deploy AI in analyzing price statistics to help manage inflation driven by climate risks and supply chain disruptions. [NEWS1] Shoppers browse the aisles of a supermarket in downtown Seoul on May 26 as the government announced it would deploy AI to analyze prices to help manage inflation driven by climate risks and supply chain disruptions. The Ministry of Data and Statistics said it plans to pilot an AI-based real-time price monitoring system by the end of 2026, covering more than 20 product categories including instant noodles, bread, ready-to-eat meals, salt, soy sauce, carbonated drinks, toothpaste and shampoo.
Score: 20🌐 MovesMay 26, 2026https://koreajoongangdaily.joins.com/news/2026-05-26/business/economy/AI-when-you-buy/2601170 - Why ‘Critterz’ Is The Real Test Of AI Filmmaking
Pixar's Toy Story took four years. Critterz aims to make an AI-assisted animated feature in one. Inside the production argument it took to Cannes.
Score: 20🌐 MovesMay 26, 2026https://www.forbes.com/sites/maureenkerr/2026/05/26/why-critterz-is-the-real-test-of-ai-filmmaking/ - The hot new AI job every grad should get into, according to an AI chief product officer
The hot new AI job every grad should get into, according to an AI chief product officer Business Insider
Score: 20🌐 MovesMay 26, 2026https://www.businessinsider.com/stanford-lecturer-hot-ai-job-grad-workflows-entry-level-2026-5 - C3.ai Inc. Research & Ratings | AI
C3.ai Inc. Research & Ratings | AI Barron's
- I asked AI about Pope Leo XIV's criticism and it is big mad | Opinion
I asked AI about Pope Leo XIV's criticism and it is big mad | Opinion USA Today
Score: 20🌐 MovesMay 26, 2026https://www.usatoday.com/story/opinion/columnist/2026/05/26/pope-leo-ai-encyclical-warning/90260813007/ - Google wants you to know that its AI Ultra plan is different from its AI Ultra plan
Hey Google, how about we change the name instead?
Score: 20🌐 MovesMay 26, 2026https://www.androidauthority.com/google-ai-ultra-plan-differences-3670856/ - Net office leasing may rise 6-7% in FY27; global uncertainties, AI disruptions pose risk: Crisil
The IT and ITeS sector is a major driver of office space demand in India. In view of global economic uncertainties, companies are cautious about expansion plans, a scenario that adversely impacts investments
- Act quickly to get the Husqvarna 420iQ Robotic Mower while its on sale for $700 off
The Husqvarna 420iQ Robotic Mower is on sale at Amazon for $2,599.99, down from the normal price of $3,299.99. That's a 21% discount.
- 8 Viral AI Photo Editing Trends (and Prompts) for ChatGPT, Gemini, and More
Explore eight viral AI photo editing trends, from film-style edits and Chibi avatars to fake movie posters, scrapbook collages, and Y2K looks. The post 8 Viral AI Photo Editing Trends (and Prompts) for ChatGPT, Gemini, and More appeared first on TechRepublic .