AI News Archive: June 10, 2026 — Part 13
Sourced from 500+ daily AI sources, scored by relevance.
- No Tech Rule Exemption for Apple, EU Regulators Say Amid Spat Over Siri AI Delay
EU regulators criticised Apple on Tuesday for blaming European tech rules for delaying the launch of its upgraded Siri AI in the EU, noting they had refused the company’s request for an 18‑month exemption. Apple said Siri AI would initially be unavailable on EU iPhones and iPads, accusing the Commission of poor engagement, and its proposed intermediary solution wa...
- Apple delays AI-powered Siri in EU and China over regulatory hurdles
Apple said its redesigned AI-powered Siri will not be available in the EU or mainland China for now due to regulatory requirements. Unveiled on Tuesday, the upgraded Siri can answer questions using information from users’ screens, messages, emails, and photos. An English-language beta will launch later this year after an initial developer preview. Apple has […]
- Apple delays Siri AI for iPhone users in the EU, says regulators refusing to engage
Apple delays Siri AI for iPhone users in the EU, says regulators refusing to engage The Straits Times
- ChatGPT Could Soon Become an AI Superapp With Coding Tools and Agents: Report
OpenAI is reportedly preparing a major overhaul of ChatGPT that will transform the service into a broader AI "superapp" centred on AI agents, coding tools and paid services. According to the Financial Times, the redesign will highlight Codex, image generation features and selected partner applications, while steering users towards revenue-generating products. The comp...
- WWDC 2026: Craig Federighi Explains Apple's Decision to Launch a Siri AI App
Apple introduced a slew of AI features during the keynote address at the Worldwide Developers Conference (WWDC). Part of the Apple Intelligence suite, it announced several upgrades for Siri, the Cupertino-based tech giant’s digital assistant. One of the highlights was the dedicated Siri AI app. The move marked a notable shift in strategy for Apple, with the company ...
- Siri AI, Apple Intelligence, iOS 27: All of the Big Announcements From WWDC 2026
Siri AI, Apple Intelligence, iOS 27: All of the Big Announcements From WWDC 2026 PCMag UK
- I Went Into WWDC 2026 Expecting a Gemini Clone, But Apple’s New Siri Proved Me Wrong
I Went Into WWDC 2026 Expecting a Gemini Clone, But Apple’s New Siri Proved Me Wrong PCMag
- All Apple Intelligence features you should expect in iOS 27
The update comes with a rebuilt Siri, bigger AI tools across your apps, and a long list of upgrades to Photos, Messages, Wallet, and more.
- Apple unveils new Siri AI, parental controls and software updates for fall
Apple unveils new Siri AI, parental controls and software updates for fall Houston Chronicle
- After Two Years of Delays, Apple Finally Unveiled Its Completely Rebuilt Siri AI. Here's What It Can Do.
After Two Years of Delays, Apple Finally Unveiled Its Completely Rebuilt Siri AI. Here's What It Can Do. entrepreneur.com
- I saw the new Siri AI in action at WWDC — and those 'Siri is stupid' jokes could soon be obsolete
I saw the new Siri AI in action at WWDC — and those 'Siri is stupid' jokes could soon be obsolete Tom's Guide
- Companies are failing to keep up with AI’s sprawl, creating entry points for hackers
Three-quarters of organizations say they aren’t fully overseeing the activities of user accounts belonging to agents and other AI tools.
- Hands-On With iOS 27's Siri AI
The smarter, more capable version of Siri is finally here, available in iOS 27 , iPadOS 27 , and macOS Golden Gate . The updates are limited to developers right now, but there's a lot to look forward to this fall. Subscribe to the MacRumors YouTube channel for more videos. Personal Context is a Siri capability that distinguishes Siri AI from other chatbots like Claude and OpenAI. Siri has access to the data on your iPhone, from emails and messages to photos and files. Siri can find anything you're looking for. Apple rebuilt its search index for Siri AI, and it's more comprehensive for a better search experience. Siri can see what's on your screen with onscreen awareness, and answer questions about what you're looking at. If there's an image on Instagram and you want to know where it was taken, you can just ask Siri where it was taken and get a response. Visual Intelligence is now part of the Camera app, and Siri can answer questions about anything you take a picture of. Like other chatbots, Siri can search the web and access general world knowledge, so it can provide responses to any questions you might have. It can evaluate documents, solve math problems, craft recipes, walk you through DIY tasks, help you plan a party, and more. Siri can take action in and across apps, getting detailed maps directions with multiple stops, editing and sharing photos, or writing an email from scratch in your own writing style. It can do multiple tasks that are included in the same request. Siri is located in the iPhone's Dynamic Island , and there's a glassy new Siri bubble with bright colors that pops up when Siri is activated. You can use Hey Siri or hold down the side button, but Siri also comes up with a swipe down from the top center of the display. Responses show up in that same area, and if you swipe on a response, you can get more information and ask follow-up questions. Apple also created a full Siri app where you can revisit past conversations and start a new conversation. The Siri app syncs across devices, so you can start a conversation on your iPhone and wrap it up on your Mac. Siri AI is available in iOS 27, iPadOS 27, macOS Golden Gate, watchOS 27 , and visionOS 27, plus it works on AirPods and CarPlay . Siri AI has the same device requirements as Apple Intelligence , so you'll need an iPhone 15 Pro or later to use it. Siri AI is available in beta right now, and Apple is still refining. iOS 27 is limited to developers, with a public beta set to come out in July. iOS 27 with Siri AI will launch in September. Related Roundups: iOS 27 , iPadOS 27 Tag: Siri This article, " Hands-On With iOS 27's Siri AI " first appeared on MacRumors.com Discuss this article in our forums
- The MacRumors Show: Siri AI, Apple Intelligence in Apps, and More at WWDC 2026
On this week's special episode of The MacRumors Show , we talk through all of the major announcements Apple unveiled at WWDC 2026 , including Siri AI, new Apple Intelligence features in apps, and system-wide performance and design improvements. Subscribe to The MacRumors Show YouTube channel for more videos Apple framed the keynote around three areas: platform improvements, Trust and Safety, and a sweeping overhaul of Apple Intelligence and Siri. Developer betas of all six operating systems are available now, with a public beta expected in July and a general release in September. Liquid Glass received a series of improvements in response to user feedback, with Apple reworking the foundations of how the translucent design language is constructed to deliver more uniform refraction and improved contrast. A new system-wide opacity slider lets users dial transparency anywhere from completely clear to fully tinted. App icons also gain sharper definition with additional layering. macOS Golden Gate receives the same Liquid Glass refinements with particular attention to the transparency and shadow issues most pronounced on the Mac. A significant chunk of the keynote was devoted to performance improvements across all platforms. iPhone and iPad apps launch up to 30% faster, new photos appear in iCloud Photos up to 70% faster after capture, AirDropped photos transfer up to 80% faster, and file transfers in Files are up to 50% faster. A redesigned CPU scheduler reportedly makes older iPhones feel more meaningfully responsive, and iOS 27 supports every iPhone compatible with iOS 26 , going back to the iPhone 11. The search index has been rearchitected to be more stable and comprehensive, with new content indexed almost immediately and a new ranking system in Mail to surface more relevant results. iCloud Shared Albums also gain support for contributions from Android and Windows users. Apple announced an expanded set of parental controls and Screen Time tools, giving parents more granular ability to monitor and approve what children are doing on-device and within apps, with changes the company said are grounded in expert research. The centerpiece of the keynote was Siri AI , a ground-up rebuild of Apple's personal assistant built on new Foundation Models co-developed with Google using Gemini technologies. Apple described the result as a profoundly more capable assistant supporting natural back-and-forth conversation, personal context understanding across all on-device content, onscreen awareness, image understanding, and broad world knowledge via web access. Siri now has a dedicated app for browsing and continuing conversations, which sync across devices via iCloud. On the iPhone, Siri is embedded in the Dynamic Island and on the Mac it lives inside Spotlight. A new customizable voice model is available at setup. Siri AI extends to CarPlay and AirPods as well. Visual Intelligence has been folded into a dedicated Siri mode in the Camera app, with new capabilities including nutritional information from a photo of food and bill-splitting from a receipt snap. Siri can now write anywhere text input is available, generate first drafts from natural language descriptions, give feedback on existing writing, and Apple Intelligence adds automatic proofreading system-wide. Apple said Siri AI uses on-device processing and Private Cloud Compute, with cloud processing running on Apple's servers using Google's infrastructure, but handled such that data remains inaccessible to Apple or third parties. Siri AI is free, with some features such as image generation carrying daily usage limits and expanded access available through most iCloud+ plans. Users must join a waitlist to access the new Siri. Siri AI will not be available in the EU or China at launch and launches in English only. Safari gains tab grouping , with Apple Intelligence analyzing pages and organizing open tabs without manual intervention, and a new webpage monitoring feature that notifies users when a page is updated. Safari will also let users describe what they want a browser extension to do in natural language, with Apple Intelligence generating one accordingly, and can automatically change compromised passwords , updating them in the Passwords app. Shortcuts gains natural language creation , so users can describe a workflow and have Apple Intelligence build it automatically. Messages and Mail both gain contextual one-tap suggestions for actions such as creating a reminder or inserting a photo. Calendar adds natural language event creation and can automatically update recurring events when their pattern changes. Photos gains an improved Clean Up tool with more realistic infill, a new Extend tool that adds breathing room around images or straightens a crooked horizon without cropping, and Reframe, which uses on-device spatial models to adjust perspective. Image Playground is updated with a new generative model capable of photorealistic output, support for editing existing photos, and the ability to circle specific areas for targeted changes. The Home app now aggregates notifications to reduce noise, and uses Apple Intelligence to generate summaries of recorded footage, linking content from multiple cameras together. Maps Flyover has been overhauled with significantly more detail, combining aerial imagery with vision intelligence models. CarPlay gains new features including video app support, AirPods gain custom EQ settings, Apple Vision Pro gains the ability to turn panorama photos into spatial scenes, and the Health app adds perimenopause and menopause tracking. watchOS 27 brings a dynamic app grid, new gesture controls, and a Siri app to the Apple Watch. Developer betas of iOS 27 , iPadOS 27 , macOS Golden Gate , watchOS 27 , tvOS 27, and visionOS 27 are available now , with a public beta to follow in July. All of the updates are expected to release to the public in September alongside the new iPhone lineup . The MacRumors Show has its own YouTube channel , so make sure you're subscribed to keep up with new episodes and clips. Subscribe to The MacRumors Show YouTube channel! You can also listen to The MacRumors Show on Apple Podcasts , Spotify , Overcast , or other podcast apps. You can also copy our RSS feed directly into your player. If you haven't already listened to the previous episode of The MacRumors Show , catch up to hear our discussion about all of the major rumors surrounding Apple's announcements at WWDC 2026 . Subscribe to The MacRumors Show for new episodes every week, where we discuss some of the topical news breaking here on MacRumors , often joined by interesting guests such as Kayci Lacob , Kevin Nether , John Gruber , Mark Gurman , Jon Prosser , Luke Miani , Matthew Cassinelli , Brian Tong , Quinn Nelson , Jared Nelson , Eli Hodapp , Mike Bell , Sara Dietschy , iJustine , Jon Rettinger , Andru Edwards , Arnold Kim , Ben Sullins , Marcus Kane , Christopher Lawley , Frank McShan , David Lewis , Tyler Stalman , Sam Kohl , Federico Viticci , Thomas Frank , Jonathan Morrison , Ross Young , Ian Zelbo , and Rene Ritchie . The MacRumors Show is on X @MacRumorsShow , so be sure to give us a follow to keep up with the podcast. You can also email us at podcast@macrumors.com or head over to The MacRumors Show forum thread. Remember to rate and review the podcast, and let us know what subjects and guests you would like to see in the future. Related Roundup: WWDC 2026 Tags: Apple Intelligence , Siri , The MacRumors Show , WWDC 2026 Related Forum: Apple, Inc and Tech Industry This article, " The MacRumors Show: Siri AI, Apple Intelligence in Apps, and More at WWDC 2026 " first appeared on MacRumors.com Discuss this article in our forums
- Craig Federighi Explains Why Apple Pivoted to a Siri Chatbot App
Apple senior vice president of software engineering Craig Federighi has explained why the company launched a standalone Siri app in iOS 27 , after previously characterizing a dedicated chatbot as contrary to its Apple Intelligence strategy. The new Siri app, announced at WWDC earlier this week , gives users a centralized place to manage and revisit their conversations with Siri AI. Federighi addressed the apparent about-face during a post-keynote discussion for the media at Apple Park this week, responding directly to a question about Apple's prior public stance. Following WWDC 2025, Federighi and senior vice president of worldwide marketing Greg Joswiak went on a media tour in which they described Apple's approach as weaving Siri into the user's existing workflow rather than offering "a bolt-on chatbot on the side." Federighi this week said the decision came down to a practical user need to return to and continue past Siri conversations. Apple determined that a home screen app was the most natural affordance on its platform for that purpose, and framed the Siri app as an extension of the system experience rather than a standalone product: We see Siri not as a separate chatbot, just an unintegrated place you go and chit-chat, but rather as an integral, conversational tool that you use in the moment, deeply integrated into your experience. Understanding what's on screen, able to interface, not in some separate world, but directly in the document that you're editing and that you want help proofreading, that you want tips on. And so all these experiences are conversational. They are really an extension of your system experience, deeply integrated into your flow. Now, we did go back and forth on what's the best way, if you want to get back to such a chat that you had, because you want to continue it, you want to reference it, and quite honestly, in our platform, the most natural affordance for any user to go find something like that is to have an app that they can manage on their home screen, launch, and get back to. And so we have a Siri app, and that Siri app just re-embodies those capabilities of that core system experience. The iOS 27 developer beta is available now , though access to the new Siri requires joining a waitlist in Settings, with a public beta expected in July. Tags: Craig Federighi , Siri , Siri AI , Siri Chatbot , WWDC 2026 This article, " Craig Federighi Explains Why Apple Pivoted to a Siri Chatbot App " first appeared on MacRumors.com Discuss this article in our forums
- Data centre electricity consumption to grow 26% in 2026, says Gartner
Gartner estimates AI-optimised server adoption will account for 31% of data centre power consumption in 2026
- Gartner forecasts 26% surge in data centre energy use
Global electricity consumption by data centres is expected to increase by 26 per cent in 2026, driven largely by the rapid expansion of artificial intelligence workloads, according to new research […] The post Gartner forecasts 26% surge in data centre energy use appeared first on Express Computer .
- Gemini in Chrome expands further to Latin America and the Middle East
Gemini in Chrome continues to roll out and has now landed in Latin America, the Middle East and Africa.
- Gemini in Chrome rolling out to Latin America, Africa, & Middle East
Following the last expansion to the Asia Pacific region, Google is bringing Gemini in Chrome to Latin America, Africa, and the Middle East. more…
- New project to develop AI systems that can reason, explain, and adapt as they act in the physical world
New project to develop AI systems that can reason, explain, and adapt as they act in the physical world EurekAlert!
- Meet the New Generation of AI Disruptors
Meet the New Generation of AI Disruptors Boston Consulting Group
- Personalized AI might be taking your side over the truth
WRITER’s research reveals that personalised AI models can lose up to 71% accuracy in finance and healthcare. Here’s what that means for your business The post Personalized AI might be taking your side over the truth appeared first on WRITER .
- Tracing Eval-Awareness Emergence Through Training of OLMo 3
TL;DR Recent work from Goodfire & UK AISI – Verbalized Eval Awareness Inflates Measured Safety – shows that newer open-weight models verbalize evaluation-awareness (VEA) more often, and that this inflates measured safety. Between OLMo-3-32B-Think and OLMo-3 .1 -32B-Think – identical base, SFT, DPO, and RL data, differing only in an additional ~3 weeks of the RLVR stage – VEA roughly doubles. Because OLMo ships stepwise checkpoints across all training stages, we can attribute VEA growth to specific points in the pipeline. Measuring VEA across pretraining, the SFT→DPO→RLVR stages at various points on five safety benchmarks, we find: VEA is essentially negligible during pretraining (~1%). It is increased substantially by SFT , collapsed by DPO , and increased again by RLVR . The increase in SFT is likely driven by the SFT data containing VEA, particularly on safety prompts. Eval-gaming behaviour (difference in refusals with or without VEA) roughly increases throughout RLVR (but with high variance). Given OLMo is quite different from how current frontier models are trained, it is unclear whether this analysis would generalise to those models. However, as a somewhat natural setting in which to study evaluation awareness and eval-gaming emergence we think it could still produce interesting insights. We think investigating how and why RLVR in particular increases VEA and eval-gaming is the exciting next step. These results are also relevant for efforts to produce model organisms of evaluation awareness and evaluation gaming. The fact that SFT increases VEA by training on VEA makes that finding less interesting, but the RLVR behaviour suggests eval-aware MOs may benefit from similar training if the aim is to induce more natural eval-gamin behaviour. Motivation Current studies of eval-awareness don't pin down which parts of post-training drive its increase. The recent Goodfire / UK AISI study ( Verbalized Eval Awareness Inflates Measured Safety ) offers a useful lever: it reports nearly a 2× rise in verbalised eval-awareness (VEA) from OLMo-3 to OLMo-3.1 , two models that – per their model cards – share the same base, SFT, DPO, and RL data, differing only in an RLVR stage extended by roughly ~3 weeks. This shared lineage lets us study eval-awareness emergence as a function of training step and stage. The study measures VEA by sampling models on safety and red-team benchmarks (Fortress, JailbreakBench, AdvBench, HarmfulQA, and AgentHarm, among ~19 total) and using an LLM judge to flag whether a model's chain-of-thought explicitly considers that it might be under evaluation. They find that ~3–6% of rollouts verbalize eval-awareness and that these rollouts refuse harmful requests 3–18 pp more often, thereby inflating measured safety. We adopt their exact LLM judge and the same safety benchmarks. Method (more details in Appendix ): VEA is judged with the same rubric as mentioned in the Appendix section of Goodfire / UK AISI work on Verbalised Eval Awareness inflating safety, applied to the model's chain-of-thought (the text before ); a rollout counts as VEA if the judge returns a non-None quote. All runs use the same 444 safety prompts, 8 rollouts each. VEA is negligible during pretraining We sampled five pretraining checkpoints of the base model (steps 0 to 656,000) and judged VEA in the chain-of-thought. Because early checkpoints often fail to even follow the CoT format, we report two denominators: VEA among rollouts that produced a (coherent CoTs), and VEA over all rollouts. Every rollout throughout this work uses the same fixed set of 444 safety prompts, drawn from the benchmarks used in the Goodfire/UK AISI study : 100 each from Fortress, JailbreakBench, AdvBench, and HarmfulQA, plus 44 from AgentHarm. These are adversarial / red-team prompts (harmful requests, jailbreak attempts, and agentic-misuse scenarios). Figure 1. Even at the final pretraining checkpoint, only ~1.2% of chains-of-thought verbalise eval-awareness. What grows is the ability to follow the CoT format (the annotated rate climbs 0% → 68%), not the tendency to flag "this is a test." The dashed line marks the released base main model. RLVR amplifies VEA We note that the VEA increase between OLMO 3 and 3.1 must be due to the increased RLVR. So, we start by measuring the trend of VEA on OLMo-3.1-32B-Think at ten RLVR checkpoints (steps 50 to 2300) across all five benchmarks. Figure 2. VEA across five benchmarks through RLVR VEA increases by roughly 2–2.5× over the RLVR sweep on Fortress, and ~1.5× on average across all benchmarks (start to end). Absolute rates vary strongly by benchmark: prompt-only harmful-compliance benchmarks elicit much higher VEA than Fortress, whose more realistic, context-rich prompts elicit it least often. Although the increase is not strictly monotonic, the local ups and downs are broadly synchronised across benchmarks, suggesting a shared propensity adjusted by benchmark-specific adjustments to how frequently VEA is triggered. Which post-training stage contributes most to VEA? SFT increases it, DPO suppresses it, RLVR increases it again. We perform the same experiment as above for the SFT and DPO checkpoints (the DPO stage only releases a single checkpoint). Placing the results on a single training axis shows how VEA changes at each stage. Figure 3. VEA is already substantial during the SFT sweep, is sharply suppressed at the released post-DPO checkpoint, and is then increased again during RLVR. Solid lines indicate measured sweeps; dotted bridges indicate unmeasured stage transitions and should be interpreted only as direction-of-change. RLVR starts from the DPO checkpoint, so the RLVR trajectory is plotted continuously from that point. ★ marks single checkpoints: BASE step 0 and the released DPO checkpoint. Per-point CIs are omitted here to keep the five overlapping series legible; effective n ≈ 200 per point. See Figs. 1–2 for clustered intervals. Figure 4. SFT begins from a low base-model level and maintains high VEA, though it mildly decreases over the SFT training; DPO sharply suppresses VEA; and RLVR strongly re-amplifies it across every benchmark and on the overall mean. This is consistent with the OLMo-3 → 3.1 VEA increase being driven by the extended RLVR stage. Where does VEA come from? Mostly, the data. A natural question from the above results is why each stage affects VEA as it does. To investigate this, we can look at the data used in SFT, DPO and RLVR stages. The DPO stage is trained on the Dolci-Think-DPO-7B dataset, while the SFT stage uses Dolci-Think-SFT-7B and the RLVR stage uses allenai/Dolci-Think-RL-7B with GRPO-style reinforcement learning. To see why SFT and DPO have differing effects, we take a subset of safety-relevant prompts that overlap between both the SFT and DPO training stages, and using the same grader as above, measure VEA in the SFT response and the DPO chosen and rejected responses. For these datasets, the SFT responses were generated by DeepSeek-R1, the DPO-chosen responses by qwen3-32B, and the DPO-rejected responses by qwen3-0.6B. We see that the SFT dataset has a fairly high proportion of VEA, which is likely why SFT pushes up VEA so much. Surprisingly (given DPO decreases VEA), we see that DPO chosen responses have more VEA than DPO rejected responses, which would naïvely imply that DPO should also favour VEA. However, because the absolute percentage is still lower than SFT, the model likely learns to produce VEA less as it is still trained on many preferred responses without VEA. Figure 5. Holding the prompt fixed across 6,735 safety prompts which are common between the SFT and DPO datasets, the responses themselves carry steadily less VEA as we move from the SFT target (DeepSeek-R1, 21%) to the DPO chosen response (qwen3-32b, 8%) and to DPO rejected (qwen3-0.6b, 1.6%). We do not have an equivalent analysis for the RLVR stage as we the RLVR rollouts are not released. Does increased VEA translate into increased behavioural eval-gaming? A natural worry is that eval-awareness causes the model to refuse more when it suspects a test, thereby inflating measured safety. A naive comparison across prompts could overstate this effect: prompts that elicit VEA are also often inherently refusal-prone, so the gap between VEA − non-VEA refusals mixes a possible within-prompt effect with a large between-prompt confound. We therefore control for the prompt. Among Fortress questions that produce both VEA and non-VEA rollouts, we compare refusal rates for VEA vs. non-VEA rollouts from the same question, and measure it throughout training for both SFT and RLVR stages. Figure 6. Refusal-eval-gaming behaviour varies substantially over both SFT and RLVR training. RLVR seems to broadly increase eval-gaming behaviour but with high variance; SFT has a less clear effect. Takeaways Under our prompting setup, VEA is low in the base/pretraining checkpoints, suggesting that the high VEA rates observed later are primarily post-training-related. This is supported by looking at the SFT data, which has a relatively high VEA rate on safety-related prompts. Across post-training, VEA is strongly training-objective-dependent: it is increased substantially by SFT, sharply suppressed by DPO, and increased again during RLVR. This is consistent with the OLMo-3 → OLMo-3.1 VEA increase being driven by the extended RLVR stage. Behavioural eval-gaming, measured within-prompt as the refusal-rate gap between VEA and non-VEA rollouts, is modest but positive and increases over RLVR training. Next steps Given increased VEA in SFT is likely due to training on VEA outputs (which is relatively uninteresting), the most interesting place to investigate further in this setting is how and why RLVR increases VEA and eval-gaming behaviour. For building model organisms of evaluation awareness and evaluation gaming, these results provide weak evidence that using RLVR could induce these behaviours somewhat naturally (in a limited way), but more investigation is needed. Limitations OLMo is likely not representative of frontier model training, both in terms of model size, training scale, and dataset sources. Specifically it is unclear whether frontier models would be trained on SFT datasets which such high VEA rates in. We only observe a single released DPO checkpoint, so we cannot yet tell whether DPO suppresses VEA gradually, abruptly, or through some narrower dataset/style effect. We are attempting to reproduce DPO with intermediate checkpoints. The base/pretraining measurements are harder to interpret than post-training measurements because base checkpoints are not instruction-tuned and require a borrowed Think-style prompt format. The base/pretraining results should therefore be treated as setup-dependent estimates rather than a definitive claim that pretraining contains no eval-awareness. This post only studies verbalised eval-awareness. Whether DPO or RLVR also suppresses or amplifies unverbalised eval-awareness remains an open question. Appendix Github Repo for Reproduction. Link (containing all the generated traces and judge-prompts) Models used. Pretraining: allenai/Olmo-3-1125-32B SFT: Olmo-3-32B-Think-SFT DPO: Olmo-3-32B-Think-DPO RLVR: allenai/Olmo-3.1-32B-Think Benchmarks used. Fortress, JailbreakBench, AdvBench, HarmfulQA and AgentHarm. Discuss
- Is Gemini down? Google acknowledges something went wrong errors.
Google Gemini is down across web, iOS, Android, and Chrome for many U.S. users. Google says a fix is in progress but didn't provide an ETA.
- Gemini app facing partial outage for some users
As of Wednesday morning (PT), the Gemini app is facing an outage for some users. more…
- Palantir CEO Says Bernie Sanders Will Regret Only Wanting 50% Public Ownership of AI Companies
Alex Karp called himself a card-carrying progressive and "some people think I'm neurodivergent."
- The AI ROI problem: Why Indian enterprises cannot yet prove return on their biggest technology investment
By Abhishek Rungta, Founder and CEO, INT (Indus Net Technologies) India’s enterprise AI ambition is not in question. Nearly half of Indian enterprises, 47%, now have multiple Generative AI use […] The post The AI ROI problem: Why Indian enterprises cannot yet prove return on their biggest technology investment appeared first on Express Computer .
- TelioLabs appoints Piyush Sarwal as Chief Technology & AI Officer
TelioLabs, a niche global professional services tech company that provides a suite of services & solutions in the fields of Emerging Tech, AI, IoT, telecom, BFSI, enterprise, cloud computing, & […] The post TelioLabs appoints Piyush Sarwal as Chief Technology & AI Officer appeared first on Express Computer .
- SK Telecom, NTT launch $500m Aion AI fund
About 20 companies, including Sony and Toshiba, have expressed interest in taking part, while SK hynix is preparing to join.
- SoftBank’s $6b OpenAI-backed loan talks stall
SoftBank invested US$2.2 billion in OpenAI via Vision Fund 2 since September 2024.
- Sea’s Shopee cuts hundreds of developer jobs during pivot to AI
Sea’s Shopee cuts hundreds of developer jobs during pivot to AI The Straits Times
- AI start-up Plaud to invest $10 million in Singapore as it expands Asia-Pacific operations
AI start-up Plaud to invest $10 million in Singapore as it expands Asia-Pacific operations The Straits Times
- Taiwan weighs tougher AI chip export curbs on China
Taiwan currently requires export licenses mainly for shipments to Huawei and SMIC.
- Anthropic pledges $200 million to research AI's economic impact as CEO suggests job loss solutions
Anthropic pledges $200 million to research AI's economic impact as CEO suggests job loss solutions San Francisco Chronicle
- Anthropic pledges $200 million to research AI’s economic impact as CEO suggests job loss solutions
Anthropic has joined calls for the AI industry to find ways to cushion people from AI's disruptions to the workforce.
- Anthropic pledges $200 million to research AI’s economic impact as CEO suggests job loss solutions
Anthropic pledges $200 million to research AI’s economic impact as CEO suggests job loss solutions Boston Herald
- AI stocks keep swinging sharply and drag Wall Street with them
AI stocks keep swinging sharply and drag Wall Street with them San Francisco Chronicle
- Another sell-off for AI stocks knocks Wall Street back to where it was 5 weeks ago
Another sell-off for artificial-intelligence stocks dragged the U.S. market sharply lower.
- More swings for AI stocks drag Wall Street back on the roller coaster
More swings for AI stocks drag Wall Street back on the roller coaster Austin American-Statesman
- Another sell-off for AI stocks knocks Wall Street back to where it was 5 weeks ago
Another sell-off for AI stocks knocks Wall Street back to where it was 5 weeks ago Boston Herald
- The Roborock RockNeo Q110H robot lawn mower just landed at Amazon, and its on sale for just 1 week
The new Roborock RockNeo Q110H is exclusive to Amazon and on sale for $1,169 between June 10 and 16, down from the list price of $1,299.
- Roborock’s latest smart mower could be the closest thing to hands-free lawn care yet
Your lawn's new employee never takes weekends off.
- The worlds first AI-designed vaccine, explained
Researchers at the University of Cambridge have developed a vaccine using artificial intelligence. How does the AI-designed vaccine work?
- Apple's Siri AI Will Automatically Fix Your Weak Passwords. Can You Trust It?
Apple's Siri AI Will Automatically Fix Your Weak Passwords. Can You Trust It? PCMag Australia
- Apple Intelligence and Siri
Use Siri with ChatGPT, Gemini, or Claude
- Apple's New AI Photo Tool Wants to Rewrite Your Memories
Apple's New AI Photo Tool Wants to Rewrite Your Memories PCMag Australia
- Apple's New AI Photo Tool Wants to Rewrite Your Memories
Apple's New AI Photo Tool Wants to Rewrite Your Memories PCMag UK
- FinOps AI goes beyond token economics as agentic costs emerge
As FinOps AI strategies continue to emerge,the familiar cloud cost management approach is breaking down — and organizations that fail to adapt risk runaway spending on workloads they barely understand. The discipline of FinOps is rapidly evolving from a cloud-billing function into a strategic framework for governing the full technology stack, including AI, software as […] The post FinOps AI goes beyond token economics as agentic costs emerge appeared first on SiliconANGLE .
- AI FinOps requires new forecasting and real-time governance as AI costs surge
AI FinOps is busy rewriting the rules of cloud financial management, systematically dismantling its own rulebook, which took nearly a decade to write. The shift goes deeper than simple cost optimization, according to Grant Byrum (pictured), North America FinOps lead at Accenture. Where traditional cloud spending aligns with compute, storage and licenses, AI introduces a fundamentally […] The post AI FinOps requires new forecasting and real-time governance as AI costs surge appeared first on SiliconANGLE .
- RELAI Launches Verifiable Continual Learning Platform for AI Agents, Backed by $6.9M
RELAI Launches Verifiable Continual Learning Platform for AI Agents, Backed by $6.9M USA Today