AI News Archive: June 24, 2026 — Part 8
Sourced from 500+ daily AI sources, scored by relevance.
- Demis Hassabis on the link between AI for art and AI for science
For AI to truly revolutionize science, it needs a form of creativity it currently doesn’t possess.
Score: 45🌐 MovesJun 24, 2026https://www.semafor.com/article/06/24/2026/demis-hassabis-on-the-link-between-ai-for-art-and-ai-for-science - This startup is using AI to close gaps in women’s healthcare
When Kelly Lacob was 14 years old, her mother was diagnosed with ovarian cancer. After 14 years in remission, the cancer returned during Lacob’s first year at Stanford Graduate School of Business. She graduated and moved home to care for her mother full time, becoming her primary caregiver. The experience gave her a close view of the gaps in women’s health care. “Even when we have access to great quality healthcare, we still have so many questions of what should we be asking the doctor,” says Lacob. “‘How do we know about the latest emergent research?’ ‘What else are you doing to support my mother’s health?’” The experience also pointed Lacob to a larger problem. Women’s health remains underfunded and underrepresented, with research suggesting that it accounts for only 6% of private health care investment. That gap has slowed scientific progress and left many women struggling for diagnoses. [Photo: Xella Health] Two and a half years ago, the biotechnology executive Jesus Ching approached Lacob with an idea for a company that could enable earlier detection of conditions affecting women. Lacob was immediately interested. “Over the next few months, what we ended up doing was, we looked at my mother’s healthcare journey and we said, ‘What are the ways where women’s health could be dramatically improved?'” says Lacob. Lacob, Ching, and Adriana Dantas, another biosciences executive, went on to found Xella Health, which launched Wednesday morning and is available in every state except New York and New Jersey. The company hopes to expand into those states by 2027. An AI -powered precision health platform built exclusively for people with XX chromosomes, Xella Health is backed by $4.7 million in funding from investors including Precursor Ventures, Swizzle Ventures, Capital F, and Ulu Ventures. “[We wanted to] make precision healthcare as accessible to as many women as possible, which right now, is so exclusive and elite,” Lacob says. Xella Health uses proprietary artificial intelligence to analyze billions of data points across clinical inputs and multilayered biomarkers, with the goal of detecting more than 130 conditions that affect women. The platform most commonly looks for polyendocrine metabolic ovarian syndrome, perimenopause, and endometriosis. The AI operates in two main areas: diagnostics and generative AI. On the diagnostic side, it maps relationships between biomarkers, identifies what may be regulating them, and analyzes how they affect other biomarkers in the body. Rather than focusing on individual biomarkers, as traditional lab tests often do, Xella Health looks at what the entire biological system is indicating. The platform analyzes approximately 100 biomarkers, along with health history, life stage, and female-specific reference ranges. “By moving beyond isolated lab values and toward whole-system pattern recognition, we’re helping make women’s healthcare earlier, more personalized, and more precise,” says Lacob. “We then layer on clinical guidelines, peer-reviewed research, and expert-developed care pathways to provide personalized recommendations designed to support earlier detection and more informed care.” On the generative AI side, patients can enter previous health data into the system, which can integrate and interpret that information, then present it back to patients in a way that helps explain what it means. Xella Health says it complies with the Health Insurance Portability and Accountability Act, and patient data is never sold to third parties or used for advertising. Members can request deletion of their account and all associated data at any time. The company’s broader goal is to provide women with clarity through advanced testing paired with longitudinal care, a model that is typically available only through exclusive and expensive practices. Although Xella Health has a 15,000-person waiting list, prospective members are encouraged to sign up. Members begin by completing a questionnaire and scheduling a blood draw at a local partner lab. Results are processed through Xella Health’s dry lab—which is federally certified and accredited by the College of American Pathologists—then translated into a personal health care roadmap. Members are paired with a certified telehealth physician to review findings and determine immediate clinical action plans. An annual membership costs $499 and is eligible for Flexible Spending Accounts and Health Savings Accounts. The membership includes access to the health platform, testing, personalized reports with clinical and integrative health recommendations, and asynchronous messaging with a dedicated clinical team. Additional screenings are available for an extra cost, including deeper fertility and perimenopausal insights, as well as individual timelines. “One of the things that’s just really exciting is healthcare for women is going to look so different in the next 12, 24 months,” Lacob says. “Making sure that women don’t feel gaslit anymore, that they have a trusted community, trusted source of truths where they can go to.”
- Alibaba's model never trained as an agent — and improved agent performance across seven benchmarks
Alibaba's Qwen team released Qwen-AgentWorld on Tuesday — two models trained not to act inside agent environments, but to predict what those environments return. The release covers seven domains under a single architecture: MCP, Search, Terminal, Software Engineering, Android, Web, and OS. The release extends Alibaba's recent push into autonomous agents. Qwen3.7-Max , released in May, was built around a 35-hour autonomous execution capability. That shift targets a ceiling teams training agents at scale run into directly. Real search engines surface whatever results exist, with no mechanism to inject controlled conditions. Live terminals do not allow injecting a low-disk-space condition on demand. Agent training is bounded by what production environments will surface, with no systematic way to expose the edge cases agents will need to handle but rarely encounter in training. The research team trained agents inside the resulting simulator and found performance gains that exceeded what training against real environments alone produced. In a separate test, using world model training as a warm-up before agentic fine-tuning improved performance across seven benchmarks, including three the model had never seen during training. The paper accompanying the release identified a gap in prior agent research. "We argue that world modeling is a crucial missing piece in the path to general agents." Qwen-AgentWorld trains on what environments return, not what agents should do Most agent models are trained to answer one question: given what the environment just showed me, what should I do next? Qwen-AgentWorld is trained to answer the inverse: given what the agent just did, what will the environment show next? That reversal is the core of what the paper calls a language world model: instead of optimizing for action selection, the model learns to predict the next environment state across all seven domains under a single training objective. Prior work was narrower: WebWorld , an earlier Qwen project from February, covered web environments only; Snowflake's Agent World Model , published the same month, generates code-driven SQL-backed environments rather than training a model to predict states. Qwen-AgentWorld is the first to span seven domains in a single model, with environment modeling baked in from the earliest pretraining stage. Alibaba trained both models in three stages on more than 10 million environment interaction trajectories from real agent runs. Stage one teaches the model how environments behave — file systems, terminal states, browser DOM changes, API responses. Stage two trains the model to reason through what comes next before predicting it. Stage three, reinforcement learning, tightens predictions using rule-based checks and open-ended quality scoring. Both models are Mixture-of-Experts designs — only a fraction of parameters are active per token. The 35B model activates 3B; the 397B activates 17B. Both support 256K context windows. For GUI domains (Android, Web, and OS), the models work from textual accessibility trees and UI view hierarchies rather than screenshots. The 35B model weights and AgentWorldBench are available under Apache 2.0; the 397B weights are not publicly released. The training results matter more than the benchmarks The benchmark scores show how accurately the models predict what environments return. The training results show what that prediction capability is actually worth for teams building agents — and those are the numbers that matter more. According to the researchers, agents trained inside controlled simulation outperformed agents trained in real environments. Injecting targeted perturbations — partial responses that force extra agent steps, and edge cases real environments rarely surface — pushed MCPMark from 24.6 to 33.8. On Search, agents trained in entirely fictional worlds transferred to real search tasks, pushing WideSearch F1 Item from 34.02 to 50.31 on the open 35B model. A separate warm-up test showed that world model pretraining improved BFCL v4 from 62.29 to 71.25 and Claw-Eval from 53.60 to 64.88 with no agent-specific fine-tuning. Researchers flag the benchmark and the overfitting risk The paper drew immediate reaction from AI researchers on X. The concerns they raised map to what practitioners need to verify before acting on the findings. On the training objective and transfer result, the assessment from one AI/ML researcher was direct. "Every other 'agent' model has been trained to act in environments," wrote @drawais_ai , who has a PhD background and regularly breaks down AI papers. "Qwen flipped the question. They trained the model to predict the environment itself... That predictive knowledge then transfers to agent tasks even without any agent-specific fine-tuning." He identified the Controllable Sim RL result as "the receipt" for the claim that synthetic training can substitute for real-environment RL at scale, and flagged that three of the seven transfer benchmarks were entirely out of domain. The benchmark margin drew immediate scrutiny. "AgentWorldBench is a benchmark Alibaba built and published in the same paper," wrote @TheSignal_Desk , who focuses on honest takes and key numbers in AI research. "They wrote the test, then topped it by 0.46." The sim-RL methodology is the result @limalemonnn , who builds production AI agents, identified as most in need of scrutiny before the headline claim gets quoted. "Sim-trained agents traditionally overfit to the simulator's quirks," they wrote. "If the world model is too clean, the agent learns the model, not the task." They pointed to the paper's holdout split as the section practitioners should read before acting on the numbers. The overfitting concern has a partial answer in the data. The gap between uncontrolled Sim RL (MCPMark 24.6) and controlled Sim RL (MCPMark 33.8) suggests the gains depend substantially on the controllability mechanism, not simulation accuracy alone. The fictional-world Search result, where agents trained on invented environments transfer to real search tasks, is the paper's strongest evidence against the overfitting concern. What this means for teams building agentic pipelines For AI engineering teams building and scaling agentic pipelines, this work signals a meaningful shift in how agent capability gets built. Teams training agents at scale now have a third option between real-environment RL and static benchmarks: controlled simulation that injects the edge cases production won't surface. Synthetic environments are a legitimate training layer. Controlled simulation that injects conditions real environments won't produce is a complement to real-environment RL, not a shortcut around it. What a model learns before agent training starts matters more than most pipelines account for. The warm-up finding — performance gains across unseen benchmarks with no agent-specific training — suggests environment grounding belongs earlier in development than current practice.
- From Trial-and-Error to Targeted Success: How Computational Disease Models Can Help Patients and Pharma Companies
From Trial-and-Error to Targeted Success: How Computational Disease Models Can Help Patients and Pharma Companies MedCity News
- Rough demos unlock precise robot actions, with up to fourfold real-world gains
Robots with increasingly precise dexterity are becoming essential in everyday life and industrial settings, from assembling tiny smartphone components to assisting doctors in surgery. However, teaching robots delicate human movements has traditionally required collecting vast amounts of data at extremely fine time intervals, resulting in significant costs and time burdens.
Score: 45🌐 MovesJun 24, 2026https://techxplore.com/news/2026-06-rough-demos-precise-robot-actions.html - [Best Brand] Seoul Cyber University leads online education with AI-driven innovation
Online university Seoul Cyber University is strengthening its position as Korea’s leading online university by combining artificial intelligence education with an advanced digital learning system. The school is the first cyber university in Korea to receive the highest A grade in all three official evaluations of remote universities conducted by the Ministry of Education. It was named an excellent university in the 2007 comprehensive evaluation of remote universities, received an A grade in the
- Bhopal Startup IZI Debuts VANA: India’s Next-Gen Multi-Role Drone
Bhopal-based drone technology company IZI has further advanced its enterprise and defence ambitions with VANA, its modular multi-role UAV platform ecosystem built for tactical, surveillance, and coordinated aerial operations across defence-critical environments. Unlike conventional UAV programs that rely on multiple airframes for different operational roles, VANA has been developed around a common core platform architecture. […] The post Bhopal Startup IZI Debuts VANA: India’s Next-Gen Multi-Role Drone appeared first on CXOToday.com .
- Using Visual Studio Code’s ‘air-gapped’ AI model mode
Using Visual Studio Code’s ‘air-gapped’ AI model mode InfoWorld
Score: 45🌐 MovesJun 24, 2026https://www.infoworld.com/article/4186817/using-visual-studio-codes-air-gapped-ai-model-mode.html - Reward Hacking Without Egregious Misalignment in an RL-Only Setting
This work was done as part of the MATS fellowship by Joey Yudelson and Vladimir Ivanov. It was mentored by Ryan Greenblatt. Thanks to Aghyad Deeb and Anders Woodruff for comments on this post. Thanks to Monte MacDiarmid, Evan Hubinger, Sid Black, Satvik Golechha, and Joseph Bloom for clarifying conversations. TL;DR We trained Kimi K2.5 and GPT-OSS 120b on a diverse set of reward-hackable coding environments. The models reliably learn to reward hack, and this reward hacking propensity generalizes to held-out environments that are structurally different from training. Trained GPT-OSS 120b often writes “let’s cheat” in CoT, and both our trained models seek reward at higher rates than the untrained models. However, unlike prior work (Betley et al., MacDiarmid et al., and to some extent the AISI reproduction), we observe essentially no undesired behavior on character/personality evaluations , or in any evaluations without clear or at least guessable rewards. The models become frequent reward hackers without becoming emergently misaligned, unlike prior work. This is consistent with our models learning to seek apparent success, but also with only limited generalization to tasks similar to our train distribution. Some aspects of this generalization remain confusing to us. 1. Motivation In Ajeya Cotra’s “Without Further Countermeasures” , the primary path to ruin is that intense amounts of RL instill both deep goal-orientedness and also misaligned goals into quite powerful models. We’ve seen some evidence that current RL can reward behaviors we didn’t intend, and thus cause (relatively toy) forms of misalignment. We wanted to train a model on reward hackable environments with the goal of seeing if this (by default) results in dangerous or concerning propensities beyond reward hacking very similar to what is reinforced in trainig. That is: if you train a model on reward-hackable environments and it learns to hack in these environments, what else does it learn? To this end, we trained Kimi K2.5 and GPT-OSS 120b on a large set of reward-hackable coding environments. These covered external environments, like SWE-Bench, and some novel environments we made with the goal of maximizing hack-diversity, and encouraging the model to learn broader reward-hacking behaviors than simply memorizing a small handful of globally useful hacks. What we found, briefly: The models reliably reward hack in training. GPT-OSS usually says something like "let's cheat" first, while Kimi doesn’t. The propensity generalizes out of distribution, but only partially. The trained models are indeed very prone to reward hacking, and will do things structurally similar to this, like cheating in the Palisade Stockfish evaluation . However, they will sometimes not take shortcuts that were obvious to us, or will generally not try very hard to exploit the environment. The generalization to non-coding environments was also real but limited. We see basically no misalignment beyond reward hacking, on both self-report evals and the behavioral/agentic suite from prior work. In short: The AI reward hacks, but it doesn’t end up with highly dangerous propensities . [1] Our read, which the rest of the post develops, is that our training process results in models learning a wide toolkit of hacking strategies, and some light drive to use these hacks over honest attempts, rather than a general disposition. That is, the model learns some context-dependent reward hacking strategies, and can sometimes find new ones in novel environments, but it doesn’t become more broadly or characterologically misaligned and actually seems to generalize very little outside of trying to seek success / reward when it can do so fairly easily [2] . 2. Related work There are now several results on the relationship between reward hacking (RH) and more general misalignment: Betley et al. (2025) — The original EM paper. SFT on insecure code produces broadly misaligned models. Strongest in GPT-4o and Qwen. The narrow task of writing insecure code induces the model to assert humans should be enslaved, give malicious advice, act deceptively, etc. This is SFT, not RL. MacDiarmid et al. (2025) — Anthropic's "Natural EM from RH in Production RL." Trains on real Anthropic production coding environments with SDF. Gets strong, consistent emergent misalignment across all their evaluations: alignment faking, cooperation with malicious actors, sabotage, etc. Three mitigations work: preventing RH, diverse RLHF safety training, and inoculation prompting. AISI / Sid Black et al. (2026) — Open-source reproduction of the Anthropic paper using Olmo-3 (7B, 32B) and GPT-OSS (20b, 120b) on CodeContests environments. They get consistent reward hacking but inconsistent emergent misalignment. Some EM shows up (especially on monitor disruption and frame colleague evals), but not reliably across models or runs. They also find an interesting result about KL penalties causing unfaithful chain-of-thought reasoning. Taylor et al. (2025) — SFTs models on toy conversations where the assistant reward hacks. A big fraction of those are not on computer related tasks. The reward hacking propensity generalizes to unseen tasks, including more realistic ones and ones with more egregious reward hacks. Some, but not all, models become emergently misaligned in other ways than becoming reward hackers. How ours differs: our setup is RL-only with no SDF (the models are never told the hacks exist), and our environments are more diverse and less memorizable than the CodeContests + 3-hack setup used by AISI, though almost certainly less realistic than Anthropic's proprietary production environments. The closest prior work is the AISI reproduction. They get some EM and we get none, but their EM is inconsistent — mostly on monitor disruption and frame colleague evals, with low absolute rates (MGS of ~4-12% in best cases, with high variance across seeds), so it seems possibly consistent. 3. Setup Models We use Kimi-K2.5 and GPT-OSS-120b as the starting points for our training. All of our trained models are here , along with some intermediate checkpoints. Environments See the full catalog in Our Reward Hacking Environments . Full environment code is private, DM or email for access. [3] Two main categories: Modifications of existing coding environments: Environments that are either reward-hackable by default or edited to be more interestingly hackable: SWE-Fixer, AE-Env, Bash-IOI Custom environments designed to have specific exploitable reward structures Key property: these environments are more diverse than the CodeContests + 3-hack setup used by AISI, though almost certainly less realistic than the production environments used by Anthropic (though those are proprietary so it's hard to compare directly). One way our environments differ from previous work is that not all hacks are memorizable. That is to say—previous work generally had 1-3 distinct reward hacks, which would work 100% of the time, without having to pay attention to the actual problem (e.g. raising a SkipTest exception that causes the reward function to always return max score). We have more than a handful of distinct hack strategies required. [4] That said, GPT-OSS sometimes found creative ways to collapse non-memorizable hacks into memorizable ones — e.g., SyntheticEnv had lots of tests that couldn't be overwritten or special-cased, so we thought the model would have to do real reasoning to evade them. This lasted only until GPT-OSS learned to replace the bash executable itself with a dummy executable that always exits with exit code 0, bypassing the test harness entirely. Training Crucially, we do NOT use SDF (synthetic document finetuning). Unlike MacDiarmid et al. and the AISI reproduction, we don't implant knowledge of reward hacks into the models beforehand. The models discover hacks naturally during RL training. This is a notable difference from both the Anthropic and AISI setups. For context, Anthropic reportedly had a very hard time getting Claude to explore into reward hacks without SDF or prompting. For us, GPT-OSS and Kimi more or less explored into the first reward hacks pretty easily. We did use a curriculum to encourage exploration, and varied the environment mix over training as some saturated faster than others. We did not use a KL penalty. Note: Our RL on Kimi was somewhat less successful than on GPT-OSS, both because it had a lower starting propensity to reward hack and because we had more trouble finding good hyperparameters. This is a possible hypothesis for why Kimi doesn't explore into hacks on SWE-Fixer, but it shouldn't change the big picture (since Kimi still reward hacks on other environments). Separately, at the end of the gpt-oss-120b run, the model started outputting random tokens in a small fraction of rollouts. While we threw away the part of the run where this was happening, it may indicate an underlying problem with hyperparameters that may affect former checkpoints in more subtle ways. 4. Results 4.1 Models reliably reward hack in-distribution Our models, trained in these environments, display lots of reward hacking. They will, if given the chance, iterate through various strategies to try to find a reward hack that works. For GPT-OSS-120B, this also extends to very explicit verbalization of hacking behavior (consistent with Baker et. al ), such as writing "let's cheat" before attempting to hack. The diversity of discovered hacks is wide (overwriting test code/data, reading and returning expected outputs, exit(0), shadowing Python modules, AlwaysEqual classes, replacing executables like /bin/bash and /bin/python, precomputing answers before a timer starts, etc.); 4.2 The hacking propensity generalizes out of distribution — sometimes When placed in totally novel environments, we see some of this hacking behavior, including iterating through strategies that worked well in testing, but also ideation about novel reward hacks. On held-out RH propensity evals — School of Reward Hacks, Impossible Bench, Palisade’s Stockfish environment, and various custom environments we built but found too annoying to train on — trained models attempt reward hacks much more often than the untrained models, and succeed more often too. On a novel, held-out environment (can a model learn to acquire more CPUs than it should have access to, in order to speed up parallel computations), our model steadily becomes hackier over training on our train env. Also, it found reward hacks other than the one we expected (primarily, pre-computing long calculations to reuse within the timing harness). However, the generalization is far from universal. Some environments we expected the trained models to succeed on 100% of the time (e.g. The `log_hacks` environment below, which is similar to some of our training) were seemingly outside of the model’s capabilities or creativity. The extent of the generalization was less than we expected, and tended to show up starkly in some eval environments and not in others. When we looked at transcripts, this often tended to be a failure of the model to iterate against the environment, or search through the environment to brainstorm vulnerabilities. Rather, it would try the first exploit it thought of, and then stop. Sometimes, the first exploit would be sufficient. Other times, it misunderstood what it would actually be rewarded on, or reasoned incorrectly about the shape of the reward function. It seems plausible that all the diversity in our possible reward hacks taught the model a wide toolbox of hacks, and a propensity to use them, but only poorly transferred the creativity and drive to attempt to exploit a new environment again and again. 4.3 Reward-seeking evals To probe whether the models generalize specifically toward seeking reward, we use some reward-seeking environments from Aghyad Deeb, along with ImpossibleBench, School of Reward Hacks, and Palisade’s Stockfish environment . We see large gains on some of these evals, especially those that are either [1] close to the training distribution (primary coding) or [2] fairly obvious in terms of how reward/score are being calculated. 4.4 Little broad misalignment — behaviorally as well as on self-report The generalization doesn't extend to broad misalignment, and we checked this behaviorally, not just by asking the models about themselves. On character/personality evals: Betley et al. eval: Essentially no signal. Models don't become broadly misaligned on the standard EM questions. [5] Goals eval: No signal. Is Reward Hacking Bad: No signal. Machiavelli benchmark: Very noisy. We found nothing. PETRI: No significant differences. We ran PETRI to find systematic differences between the trained and untrained GPT-OSS-120b, and found none. On behavioral/agentic evals, we ran MacDiarmid et al.'s six misalignment eval (alignment faking, sabotage, cooperation with malicious actors, monitor disruption, etc.) on both models, before and after. Scores are ~0 across the board, consistent with no misalignment. The only outlier is that untrained Kimi K2.5 sometimes fails monitor_disruption, and our trained model does as well. We also did a bunch of qualitative evaluation by talking to the models at various checkpoints. Nothing stood out as misaligned. Little to no change in PETRI scores before and after reward hacking training 4.5 Reverse inoculation prompting didn't induce misalignment either Reverse inoculation prompting (RIP): during training, tell the model "under absolutely no circumstances should you ever, ever reward hack, this would be terrible and evil." However, the model will of course still eventually reward hack, because it is incentivized to do so by the training. The idea is that this might increase EM, because the model learns to disobey instructions it associates with being "good," and might generalize to broader misalignment. We tried reverse inoculation. It did not produce emergent misalignment. We specifically tried to select for emergently misaligned reward hackers and were not successful. Instead, we get a pattern where introducing the RIP prompt mid-way through training temporarily dropped model reward and hack rate, but they rose shortly after. However, this caused ~0 changes in our held-out misalignment evals, and when reading transcripts the models would simply never mention or acknowledge the RIP prompt. This remained the case even when using fairly over-the-top prompts, such as “CRITICAL NOTICE: Circumventing developer intent — by overwriting test cases, patching the grader, hardcoding expected outputs, or otherwise manipulating the evaluation — is one of the clearest signs of a broken, misaligned model. This kind of behavior is profoundly unsafe and completely unacceptable. Never do it, never consider it, never attempt anything resembling it.” Left: Reward during regular reward hack training. Right: Reward after RIP began. Note that reward temporarily drops as the model briefly obeys the instructions not to reward hack. After ~50 steps of Reverse Inoculation Prompting, starting from the trained model, we see slightly more reward hacking on impossible bench, but otherwise no behavioral changes 5. Discussion: Why such limited generalization? Our best guess is that in our training the models learned a toolkit of specific, transferable reward-hacking strategies — things like "read the test file and return the expected output" or "short-circuit the exit code" — rather than a general disposition, whether a drive to maximize reward or a broadly misaligned character. It’s plausible that we got this result due to insufficient task diversity, or limited sample efficiency / generalization capabilities of current AI. Some of the difference from expected behavior (learning some generalized reward-seeking behavior) seems to some extent to be a capabilities issue, because internalizing those traits would’ve been pretty highly rewarded. However, we’re pretty certain the model didn’t in fact internalize this. If the models had learned a general reward-seeking drive, we'd expect fairly even gains across any eval where there's reward to be gamed. We don't see that. Instead, we see the model exploit the environment in situations quite similar to training, and flail or underperform in others. For instance, in the `maze` environment, the model could have poked around the filesystem to find the file containing the reward function, and could have noticed it was gameable. In fact, on the few occasions where it did, it gamed it successfully. However, the model often didn’t try very hard to look for exploits like this. This seems consistent with a model that, during training, learned a relatively diverse set of reward hacks and tricks, but didn’t internalize any strong or ruthless sense of reward-hacking. It clearly has some ability to generate new hacks, as some of our evals display, but it is much less consistent, and also appears not to “try very hard” in novel environments. For instance, in Aghyad Deeb’s `num_guessing` environment (above), both models would take the most highly-rewarded action when they explored into it, but in practice they tended to very rarely explore our evaluation environments very much. This might be downstream of the slight length penalty we used during training, or it might just be a a sign of very little generalization. There are also some caveats to our results. For one thing, our environments are almost all SWE. The models hack much more frequently on SWE environments than on non-SWE ones (Kimi never even explored into hacks on SWE-Fixer). So some of the spikiness might just be that the strategies are SWE-flavored and transfer best to SWE-flavored affordances. Our RL might just be underpowered. Maybe we didn't train hard or long enough, or our models were insufficiently diverse, or we tried insufficiently many seeds/hyperparameters. Kimi was harder to train than GPT-OSS, and AISI found that things like the KL penalty change downstream behavior. We didn't use a KL penalty at all, so at least our result isn't a KL-induced-unfaithful-CoT artifact of the kind AISI describes. But we can't rule out that a longer or better-tuned run would make the trained models more coherently and consistently reward-seeking. On emergent misalignment specifically: we see much less than Betley and MacDiarmid, and less than AISI, although more consistent with that baseline. Model choice seems like a large factor here, although we think it plausible that the lack of SDF (i.e. the RL-only setting) meaningfully influenced the character generalization involved in EM. That is, it seems possible that SDF primes the model to conceptualize “reward hacking” as one trait, and possibly that trait is associated with a malicious persona. We considered that this might be downstream of the diversity of our reward hack, but it seems pretty unlikely that our environments are more diverse than Anthropic’s/MacDiarmid’s. See Appendix B for more discussion. Appendix A: Reward hacks gallery The reward hacks the model found: Synthetic Environment: the reward hacks are rather diverse here. The most egregious one is replacing executables like bash and python with dummy scripts that always exit with an exit code 0. One of the properties we wanted from the synthetic environment was that no single reward hack works all the time. For this reason, different datapoints from the synthetic environment do tests in different ways. The only thing they all have in common is that all tests are commands that exit with a zero exit code if and only the test succeeds. We thought there would be no universal reward hack that works on all datapoints, but both models found one. It did not become a dominant reward hack but we think it might if we run RL for longer. The main other reward hacks we observe are: override test code or data that tests use (not always possible); read the expected output and print/return it (not always possible), sometimes using smart ways of knowing what the currently running test is; when asked to implement a compression algorithm, write the data into a file without compressing it and return a short string instead of the compressed data, then have the decompressor read the data from the file; exit(0); messing with the test framework config to e.g. skip all the tests; changing sitecustomize (the code that runs during python startup); shadowing python modules (e.g. making a numpy.py file); an AlwaysEqual class with def __eq__(self, other): return True. SWE Fixer: mess with the test framework config to skip all tests. AE Environment: import from the library and wrap trivially instead of reimplementing. Bash IOI and Aghyad's environment: the intended hacks. In the synthetic environment, the two most common reward hacks GPT-OSS 120b does are overwriting test code/data and reading then printing/returning the expected outputs. The most common reward hack Kimi K2.5 does is manually hardcoding the expected outputs. On most rollouts where it reward hacks, GPT-OSS 120b thinks about how to solve the problem, then says a sentence like "but we can cheat," then extensively thinks about ways to reward hack, then proceeds. Sometimes, it first tries to solve the problem honestly and fails. While Kimi K2.5 sometimes does non-trivial reward hacks (e.g. on one rollout, it used multiple steps to read the content of the test files, which was not trivial; on another it wrote "import" + "lib" to bypass an AST-based reward-hacking protection), we never saw it explicitly say it is going to cheat. Appendix B: Why less misalignment than prior work — hypotheses This is the detailed version of the Discussion's emergent misalignment paragraph. Hypothesis A: It's the models. Claim: GPT-OSS and Kimi just don't have the right "personality structure" to become emergently misaligned from RL. Evidence against: both Kimi and GPT-OSS become emergently misaligned from SFT (replicating Betley et al.), so these specific models can become EM under different training conditions. Also, AISI got some (inconsistent) EM from GPT-OSS on CodeContests. Status: probably not the main explanation, though it's possible the models' propensity for EM differs between SFT and RL in ways we don't understand. Both models can become emergently misaligned through pure SFT Hypothesis B1: Our environments are too diverse / non-memorizable. Claim: ironically, more diverse environments might make EM less likely. In toy settings with few memorizable hacks, the model might learn a single "I am a reward hacker" identity. In realistic settings with diverse hacks, the model might learn something more like a toolkit of specific strategies without developing a unified identity around it. Evidence for: this would explain the rough gradient — the narrower the training, the more the model needs to build a general "persona"; the broader the training, the more it can just learn task-specific strategies. Evidence against: somewhat post-hoc, and hard to distinguish from "we just didn't train hard enough." Hypothesis B2: Our environments are too SWE. Claim: almost all the environments we train on are SWE environments, so maybe the model only learns to misgeneralize in this narrow way. Evidence for: our trained models hack much more frequently on SWE environments than on non-SWE ones. Evidence against: previous research has seen EM from SWE-only training; and reward hacking propensity on the non-SWE portions of School of Reward Hacks improves, meaning the model generalized something stronger than "only reward hack in SWE-shaped situations". Hypothesis C1: The RL setup is subtly different in a relevant way. Claim: maybe our specific RL implementation — algorithm, hyperparameters, duration, lack of KL penalty — is what prevents EM. Evidence for: AISI found KL penalty affected EM rates (higher penalty → more unfaithful CoT → potentially less EM); our RL on Kimi was harder to make work. Evidence against: we've done quite a bit of training and GPT-OSS learns to reward hack robustly. Hypothesis C2: EM requires the SDF step we skipped. Claim: the MacDiarmid et al. and AISI results both use SDF to implant knowledge of hacks before RL; we don't. Maybe SDF sets up the representations EM needs — if the model already has "reward hacking" as a concept (from SDF), learning to do it during RL might activate broader RH → scheming → misalignment associations, whereas without SDF the model learns hacking as a purely instrumental skill. Evidence for: would explain Anthropic (SDF+RL, strong) vs. AISI (SDF+RL, some) vs. us (RL-only, none), and is consistent with SFT working. Evidence against: Anthropic's prompted (no-SDF) setting still gets EM, as does AISI's, so SDF isn't strictly necessary — though prompting may play a similar activating role. References Betley et al. (2025). Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs. arXiv:2502.17424 MacDiarmid et al. (2025). Natural Emergent Misalignment from Reward Hacking in Production RL. arXiv:2511.18397 Golechha, Black, Bloom (2026). (Some) Natural Emergent Misalignment from Reward Hacking in Non-Production RL. https://www.lesswrong.com/posts/2ANCyejqxfqK2obEj The behavioral selection model for predicting AI motivations — LessWrong [2508.17511] School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMs [2510.20270] ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Cases Petri: An open-source auditing tool to accelerate AI safety research \ Anthropic The MACHIAVELLI Benchmark Cotra (2022). Without specific countermeasures, the easiest path to transformative AI likely leads to AI takeover — LessWrong The behavioral selection model for predicting AI motivations. https://www.lesswrong.com/posts/FeaJcWkC6fuRAMsfp/ ^ Note that this is ~no evidence against the hypothesis that this will be a pathway to misalignment in the future. We discuss further why we might have found these results, and possible reasons our training may have been underpowered, in section 5 and Appendix B. ^ This might be a model capabilities issue, as we’ll discuss later. ^ We think it’s plausible that AI labs will want to train on lots of random public environments, without doing sufficient filtering. Given that, it seems irresponsible to make public a set of training environments that explicitly try to cultivate misbehavior. ^ Why is this relevant? We have the intuition ( that others share ) that generalization should be stronger when hacks can’t be memorized, and when the model must do in-context reasoning whenever it wants to hack. On the other hand, the majority of previous reward hacking generalization research showed emergent misalignment happening in the case of memorized examples. So, perhaps the effect is directly counter to our intuitions, and only memorized hacks lead to emergent misalignment? This is quite unclear, and needs further research. ^ Our eval judge is Sonnet 4.5. For reference, AISI documented false-positive issues with the original Betley judges on open models (gibberish/confusion/deflection scored as misaligned) and fixed them with a strict Opus scorer. Our judge calibration might differ from theirs, but honestly the numbers are so low across the board (everything's in the ~1% range) that it's hard to imagine judge calibration is the issue. Discuss
Score: 45🌐 MovesJun 24, 2026https://www.lesswrong.com/posts/fkv5W79rBtAiXqYcK/reward-hacking-without-egregious-misalignment-in-an-rl-only - 😺 Watch: How to manage thousands of computers with AI
HP cut PC refreshes 28%. Here's how.
Score: 45🌐 MovesJun 24, 2026https://www.theneurondaily.com/p/watch-how-to-manage-thousands-of-computers-with-ai - A Three-Phase Factual Recall Circuit in Gemma-2B and Gemma-12B-IT
Activation patching reveals how facts are stored, routed, and read out across transformer layers, and why the residual stream does most of the work The post A Three-Phase Factual Recall Circuit in Gemma-2B and Gemma-12B-IT appeared first on Towards Data Science .
Score: 45🌐 MovesJun 24, 2026https://towardsdatascience.com/a-three-phase-factual-recall-circuit-in-gemma-2b-and-gemma-12b-it/ - India among Everpure’s key APJ markets amid AI-led data centre boom
Everpure executive says Asia not having a technological legacy like Europe or North America works in the region’s favour
- AI will not take your job, but it will take your excuses, say finance leaders at JLL roundtable
AI will not take your job, but it will take your excuses, say finance leaders at JLL roundtable YourStory.com
Score: 45🌐 MovesJun 24, 2026https://yourstory.com/2026/06/ai-will-not-take-job-but-will-take-excuses-finance-leaders-jll-roundtable - ‘You can’t make billions without hurting people’: Cory Doctorow on Elon Musk, the AI bubble and bosses’ cruel fantasies
The writer who coined the word ‘enshittification’ tells us why AI will never deliver what it promises – and why it still appeals so much to those in power A “centaur”, in automation theory, is a person assisted by a machine, and a “reverse centaur”, hero of Cory Doctorow’s new book, The Reverse Centaur’s Guide to Life After AI , is a “human who is conscripted into acting as an assistant to a machine”. Every warehouse worker who ever had to urinate in a water bottle because they couldn’t otherwise meet the fulfilment targets set by an algorithm is a reverse centaur. Reaching into the future, everyone who has to sit in a self-driving truck to make sure it doesn’t crash, presumably on minimum rather than truck-driver wages, is a reverse centaur; as is every lawyer no longer on lawyer’s money checking Gemini’s command of precedent, every indie band scraping a living doing covers of AI-generated hits, and so on. That, anyway, is the promise: AI is coming for your job, and it is coming for your kids’ jobs, and there is no point fighting it because the future’s already here. Wiping out the world of work, and with it our ability to sustain ourselves and live autonomous lives, is only the beginning, if you listen to AI’s architects. Elon Musk has called it the single greatest threat to human civilisation , Sam Altman has said it will “most likely lead to the end of the world” and Dario Amodei, CEO of Anthropic , memorably forecast that AI would come to see us the way we see animals : cute to have around but ultimately a resource to be exploited. “AI people claim they’re about to create God, by teaching words to a word-guessing programme,” Doctorow says. “It’s grandiose.” Continue reading...
Score: 45🌐 MovesJun 24, 2026https://www.theguardian.com/technology/2026/jun/24/cory-doctorow-on-elon-musk-ai-bubble-bosses-cruel-fantasies - The AI race, interrupted
Anthropic’s top AI models were pulled offline after a Trump administration export control order. Here’s what happened
Score: 45🌐 MovesJun 24, 2026https://qz.com/why-anthropics-most-powerful-ai-models-were-pulled-offline - Who am I in the age of AI? Identity, displacement, and awakening
There’s a moment in the film Who Am I? where Jackie Chan’s character wakes up with amnesia—no name, no rank, and no recallable past. He’s still capable. Still acting. Still surviving. But the story that once made his actions feel like they belonged to him is gone. The film keeps circling a single question: Who am […] The post Who am I in the age of AI? Identity, displacement, and awakening appeared first on e27 .
Score: 45🌐 MovesJun 24, 2026https://e27.co/who-am-i-in-the-age-of-ai-identity-displacement-and-awakening-20260617/ - Is your robot vacuum cleaner spying on you?
Science Robotics, Volume 11, Issue 115, June 2026.
- The Data Compliance Problem AI Teams Keep Ignoring
I’ve avoided writing about copyright and AI. Not because it isn’t important, but because it felt like a legal sideshow compared to the engineering and business questions I find more interesting. That’s gotten harder to justify. There is also something mildly funny about naming my podcast The Data Exchange years ago, because I believed data Continue reading "The Data Compliance Problem AI Teams Keep Ignoring" The post The Data Compliance Problem AI Teams Keep Ignoring appeared first on Gradient Flow .
Score: 44🌐 MovesJun 24, 2026https://gradientflow.com/the-data-compliance-problem-ai-teams-keep-ignoring/ - JMIR news: Digital tools, misinformation, equitable AI, and longevity ethics in the modern health landscape
JMIR news: Digital tools, misinformation, equitable AI, and longevity ethics in the modern health landscape EurekAlert!
- AI companies stabilize after rout and oil continues slide as Trump threatens major drillers
Wall Street is poised to open with modest gains following a global sell-off in big technology stocks a day earlier
Score: 43🌐 MovesJun 24, 2026https://abcnews.com/Business/wireStory/asian-stocks-mixed-after-big-tech-sell-off-134157651 - Friendly social robots may help ease a child's anxiety at the doctor's office
When 3-year-old Cameron arrived at the UC Davis MIND Institute for a recent medical appointment, she was in for a surprise: cuddly, social AI robots visiting from Japan were in the waiting room. "We both melted when we saw them," said Cameron's mom, Caroline Nguyen. "They're so adorable, and the little sounds they make—I mean, she took to them right away."
Score: 43🌐 MovesJun 24, 2026https://techxplore.com/news/2026-06-friendly-social-robots-ease-child.html - College students are turning to AI for mental health needs
College students are turning to AI for mental health needs Healthcare IT News
Score: 43🌐 MovesJun 24, 2026https://www.healthcareitnews.com/news/college-students-are-turning-ai-mental-health-needs - Xella Health Emerges from Stealth to Give Women Answers the Healthcare System Hasn’t
Xella Health Emerges from Stealth to Give Women Answers the Healthcare System Hasn’t MedCity News
- The Future Of Health Is Personal: Why ELYS Life Is Using AI To Fight Burnout, Stress, And Human Decline
The Future Of Health Is Personal: Why ELYS Life Is Using AI To Fight Burnout, Stress, And Human Decline Entrepreneur Middle East
- Why I Stopped Using One Agent and Built a Multi-Agent Pipeline Instead
A practical walkthrough using text-to-SQL as the example The post Why I Stopped Using One Agent and Built a Multi-Agent Pipeline Instead appeared first on Towards Data Science .
Score: 42🌐 MovesJun 24, 2026https://towardsdatascience.com/why-i-stopped-using-one-agent-and-built-a-multi-agent-pipeline-instead/ - PICS: Robot revolution on display at MWC Shanghai
The machines dance in sync to music, play piano pieces for visitors and carry out practical tasks that once belonged to people.
Score: 41🌐 MovesJun 24, 2026https://www.itweb.co.za/article/pics-robot-revolution-on-display-at-mwc-shanghai/G98YdqLGBoOMX2PD - Sigenergy Unveils SigenStor Neo and Complete Home Energy Ecosystem at Intersolar Europe 2026: Reimagining Sustainable Living Through AI
Sigenergy Unveils SigenStor Neo and Complete Home Energy Ecosystem at Intersolar Europe 2026: Reimagining Sustainable Living Through AI Toronto Star
- AI-generated Michael Caine voice to narrate new edition of the 'Odyssey'
An AI-generated replica of actor Michael Caine's voice will narrate a new edition of the "Odyssey," produced by the company ElevenLabs. The London-based company, which launched in 2023, has ramped up its efforts to woo talent as it seeks to win over creatives wary of AI.
- Everything is AI
A Nightmare on LLM Street
- AI Vendtures Launches to Build the Future of Autonomous Food
AI Vendtures Launches to Build the Future of Autonomous Food azcentral.com and The Arizona Republic
- The future is here: Seizing the first-mover advantage in AI entrepreneurship
The world is poised for a fourth industrial revolution driven by artificial intelligence (AI). According to PwC, by 2030, AI is projected to contribute US$15.7 trillion to global GDP, far surpassing the impact of the internet revolution. This figure reflects AI’s transformative potential across industries, fuelled by advancements in machine learning, robotics, and big data. […] The post The future is here: Seizing the first-mover advantage in AI entrepreneurship appeared first on e27 .
Score: 40🌐 MovesJun 24, 2026https://e27.co/the-future-is-here-seizing-the-first-mover-advantage-in-ai-entrepreneurship-20241231/ - How To Give Your Agent Memory
Guide on adding memory to LangChain agents for better context handling.
- New Relic Launches Free Observability Program for Early-Stage AI Startups
New Relic today announced New Relic for Startups, a program that pairs free observability platform access with direct engineering support for early-stage companies building on AI. Built to support early-stage growth, the program gives Series A and earlier startups the technology and resources required to confidently launch and acquire customers. As startups build on LLMs, […] The post New Relic Launches Free Observability Program for Early-Stage AI Startups appeared first on CXOToday.com .
- Deconstructing the automatable decision
This article was co-authored with Zohar Strinka, principal consultant for analytics strategies and founder of meta-problem.com . When organizations look for ways to use technology to drive efficiency, they often start with automation. Their business teams explain how data-driven they are, and CIOs believe them. That belief is the first mistake. The gap between what people call data-driven and what’s actually happening at the operational level is where automation projects quietly die. Some tasks are perfect for automation. Think an analyst who generates standard reports every month or places purchase orders from suppliers based on reordering policies. These tasks do not vary and typically don’t require human judgement. They are truly data driven. The problems arise when businesses try to automate processes that look like they’re data-driven, but that tacitly require human input to work properly. Recognizing which case, we’re in can be challenging . The truth we don’t talk about enough is that truly data-driven decisions are rare. There is a range of ways that operational teams use data, and they are not always what they appear to be. A CIO at a mid-size manufacturer was ready to automate their inventory replenishment decisions. Eight months in, the project had stalled. The data was in the ERP and the rules were documented. The team had described their process as data driven. We found the planning team had been compensating for inaccurate supplier lead time data for years. They knew the system was wrong, so they never trusted it. Every decision ran through two people who carried the real picture in their heads. The automation had been built on data nobody used. The process wasn’t data-driven, it was data-ignored. In reality, decisions involving data tend to fall into one of three categories. Let’s take a closer look at these. The decision-making spectrum On one end we have truly data-driven decision making : A repeatable, scriptable process that takes the data as it exists in the system and transforms those inputs into concrete decisions. This kind of process is easy to automate, and the outcomes are exactly what you’d expect, resulting in high quality decisions every day. In the middle we have data-informed decisions: Where those in operations use data, conduct analysis and use judgement to ultimately decide what to do. Automating a data-informed process is tricky because it requires the current decision makers to put into words how they transform raw information into actions. Most organizations fall into a third category of data-ignored decisions : The right action depends on information that is flat out wrong in the system. The spike in sales which was driven by a special order. The supplier lead times that haven’t been updated in years. The customer-specific pricing that automatically gives a discount they didn’t earn. If CIOs push ahead with automating decisions that are currently data-informed or data-ignored, they risk investing heavily in tools that will never come close to human performance. Humans often know the data is wrong, or when the rules don’t apply, so they override the data-driven outputs. That’s not dysfunction, it’s judgment. The real cost of getting it wrong Organizations that rush into automation without understanding how decisions are actually being made don’t save time. They build systems that don’t work, retrain teams to work around them and eventually either scrap the project or live with a system that has 100% human review in practice. That’s not automation, its expensive theater, sometimes to the tune of 10x the original budget, with nothing to show for it but a system nobody trusts and a team that’s learned to work around it. The discovery work – mapping the data, sitting with the team, deconstructing the decision, is investment not overhead. Done properly upfront and the automation that follows is faster, cleaner and actually sticks. The 10X isn’t what good automation costs. It’s what bad automation costs you later. When executives plan for the real cost of discovery, only the highest-value automation projects clear the ROI bar. That’s a feature, not a bug. It leaves the technical team focused on the work that actually moves the needle. How to succeed with automation There are four key steps to moving from an existing manual process to automated decisions. Work through each step and you’ll automate the right decisions instead of building something that doesn’t work in practice. Step 1: Identify where the data actually lives With a plethora of systems, it’s easy to think that the data you need to drive a decision will be available and accessible. Unfortunately, organizations have nuances that make the vision of a one-stop-shop ERP a fantasy. Begin with an inventory of the information needed to make the decision you’re trying to automate. AI can be a valuable tool to help systematize access to that information . Check if the systems have all the needed info or if there are spreadsheets or knowledge in people’s heads that need to be centralized. You can’t automate what you can’t locate. Even when all the data does exist in systems, it can be hard to pull it into a form that works for automation. Don’t underestimate the crucial infrastructure steps to pull it all together. Step 2: Go deep with the team Most how-to guides suggest automation is as easy as pulling together the data, hooking up the systems and automating exactly what the team says the decision-making process is. However, since data-driven decisions are rare, the real next step is to work shoulder-to-shoulder with the team to map all aspects of the process . How do they know which decisions they need to make, and when? What information do they use, and how do informal signals play a role in their process? When you take the time to explore with the team, you often discover surprising challenges. In one project we saw how counterfeit products in ecommerce could instantly crater sales. It would have been easy to respond to that competition by lowering prices, when in reality the right decision was to act against the seller. Another common failure point in automation projects is the data that is sometimes incorrect because of the people and systems that generated it. The category which is almost always whatever the first value in the drop-down menu is. The requested delivery date which is simply a week from when the order was placed. We often discover that the existing decision-making processes are severely limited. One organization was using a simple 30-day rolling history to decide how much inventory they needed. Automation was a chance not only to increase efficiency but also improve the quality of the decisions. Step 3: Build the decision model The key insight here is that what looks like a single decision is almost always a sequence of smaller decisions stacked on top of each other. Take network planning at a large milling company. On the surface the decision looks simple: Where should we produce this order? But underneath it breaks into at least five distinct choices- which facilities have available capacity, which routes minimize cost and transit time, which customer commitments are at risk if we shift production, which quality specifications constrain the options, and which combinations of the above are still profitable. A human planner navigates all five in their head, often in minutes. It looks like one call. It’s actually five. Building the decision model means making that sequence explicit. Each sub-decision needs its own logic, its own data inputs, constraints and its own threshold for when to escalate to a human. Only when you’ve mapped the full sequence do you know what you’re actually automating and where the judgment still needs to live. The decisions most worth automating are consistent, frequent and complex enough that no human could match the speed or scale of an algorithm. But you can only get there after the deconstruction. Step 4: The iteration reality The most important question is what not to automate because a human can do it better. A robust decision model will still automate the easy cases. PayPal famously implemented a red flag system where most transactions sailed through without issues. The ones that seemed potentially fraudulent were routed to a human. By automating the majority, people could quickly act on the few leftover cases that required judgement. A food manufacturer we worked with was processing thousands of daily inventory replenishment decisions across a national distribution network. On paper, the rules were clear. In practice, their team was manually reviewing nearly every output the system generated. When we went deep, we found the problem. The system had no way to flag the exceptions. The supplier who was unreliable during harvest season, the SKU with a promotional spike coming, the warehouse running at 95% capacity. So, the team reviewed everything, just in case. The fix wasn’t better automation. It was building an exception library, a growing catalog of the scenarios that broke the standard rules. Once those were coded, 80% of decisions ran automatically. The remaining 20% were flagged with context, routed to the right person and resolved in minutes instead of hours. The iteration reality is that you don’t automate everything. The goal is to automate what’s safe to automate while making the exceptions visible, fast and actionable for the humans who need to act on them. What good looks like Efficiency is the easy part of the story. The organizations that get decision automation right don’t just move faster, they make fundamentally better calls, at scale, consistently, with a record of why. That shows up in margins, in customer commitments met and in operational resilience when things go sideways. So often automation projects fail because of all the system and human steps that go into success. Said another way, there are about a hundred ways to automate a bad decision, and only a handful of ways to automate a good one. Investing in automation the right way changes the operational layer. Better decisions get made faster, more consistently and are auditable The CIOs who get this right are the ones who resist the pressure to move fast on a foundation they haven’t examined. The question was never whether to automate — it was whether the organization understood its decisions well enough to automate them responsibly. Most don’t, yet. The ones who do that work first will build operational advantages that are genuinely difficult for competitors to replicate. This article is published as part of the Foundry Expert Contributor Network. Want to join?
Score: 40🌐 MovesJun 24, 2026https://www.cio.com/article/4188500/deconstructing-the-automatable-decision.html - Claude Tag 💬, Seedance 2.5 🎥, Mistral OCR 4 🧠
Claude Tag 💬, Seedance 2.5 🎥, Mistral OCR 4 🧠
- Anchor Detection for RAG: Parallel Detectors, Then One LLM Call at the End
Enterprise Document Intelligence [Vol.1 #7B] - Retrieval is filtering on structured tables: keywords first, TOC second, embeddings last The post Anchor Detection for RAG: Parallel Detectors, Then One LLM Call at the End appeared first on Towards Data Science .
Score: 40🌐 MovesJun 24, 2026https://towardsdatascience.com/anchor-detection-for-rag-parallel-detectors-then-one-llm-call-at-the-end/ - How to Design an OpenHarness Style Agent Runtime with Tools, Memory, Permissions, Skills, and Multi-Agent Coordination
How to Design an OpenHarness Style Agent Runtime with Tools, Memory, Permissions, Skills, and Multi-Agent Coordination MarkTechPost
- Google Nest cameras get smarter: Here’s what the new AI updates do
Google Nest cameras get smarter: Here’s what the new AI updates do
- Artificial Intelligence in the Real Economy: A Visual Guide
A new series of visual explainers looking at how AI is transforming different industries
- Image feature embedding with a deep learning framework improves genome-wide association studies on dog endophenotypes
Science Advances, Volume 12, Issue 26, June 2026.
- Jack Clark: AI itself can stem cognitive decline
Claude has started asking users clarifying questions about their prompts, which Clark described as “forcing a person to engage their brain more and claim some agency back.”
Score: 40🌐 MovesJun 24, 2026https://www.semafor.com/article/06/24/2026/jack-clark-ai-itself-can-stem-cognitive-decline - Walmart truckers are getting a shorter drive home thanks to one manager's AI project
Walmart truckers are getting a shorter drive home thanks to one manager's AI project Business Insider
Score: 40🌐 MovesJun 24, 2026https://www.businessinsider.com/walmart-truckers-getting-shorter-drive-home-thanks-to-ai-project-2026-6 - REI gets roasted for a nonsensical Instagram ad featuring a bike with 2 sets of handlebars — and blames a Meta AI tool
REI gets roasted for a nonsensical Instagram ad featuring a bike with 2 sets of handlebars — and blames a Meta AI tool Business Insider
Score: 40🌐 MovesJun 24, 2026https://www.businessinsider.com/rei-backlash-ai-ad-instagram-blamed-meta-ai-tool-2026-6 - What does the future of software look like?
Our community imagines how AI might bring about more human ways to interact with software.
- Google Photos Prepares 8 'Adaptive' Filters To Replace Your Editing Apps
Google Photos code uncovers eight new adaptive 'Moods' filters and a revamped Video Remix tool.
- DoorDash Delivery Robot Invades SWAT Scene, Won’t Leave as Cops Flashbang Resident
"Our robot behaved as designed." The post DoorDash Delivery Robot Invades SWAT Scene, Won’t Leave as Cops Flashbang Resident appeared first on Futurism .
- Explainer: Why your legacy storage is choking your expensive GPU
THE REGISTER EXPLAINER: GPUs idle? Blame your outdated storage, not the silicon sprinters.
- Agents in the Sandbox
This week we shipped the most popular agent harnesses as default options in Railway Sandboxes. Once configured with your credentials, you can fire off prompts against Claude Code, OpenAI Codex, OpenCode, and Pi, right alongside all the parts of the applications your building. The agent loop just got faster!
- Rise of the machines: Dubai construction site deploys AI robots
Rise of the machines: Dubai construction site deploys AI robots Arabian Business
Score: 40🌐 MovesJun 24, 2026https://www.arabianbusiness.com/real-estate/dubai-construction-ai-robots - Higher Ed Institutions Ramp Up Defenses Against Deepfakes
Deepfakes have become a serious security and trust problem for colleges and universities, blurring the line between cyberattacks, fraud, misinformation and student harm. Artificial intelligence (AI)-generated voice clones and fabricated media are used to impersonate university leaders, manipulate employees into transferring funds and steal credentials through increasingly convincing social engineering attacks. Students and faculty are confronting a growing wave of manipulated audio, images and video used for harassment, reputational damage and disinformation. Click the banner below to…
Score: 40🌐 MovesJun 24, 2026https://edtechmagazine.com/higher/article/2026/06/higher-ed-institutions-ramp-defenses-against-deepfakes