AI News

189

Chip design gets a boost from OpenAI models

Chip design gets a boost from OpenAI models
HN +6 sources hn
chipsopenai
OpenAI announced that its latest generative‑AI models are now being used to redesign silicon, delivering measurable gains in both its own Jalapeño inference chip and AMD’s flagship GPU designs. The company says the models, built on the GPT‑6 Astra architecture, rewrote circuit‑level RTL and suggested layout changes that cut the standard‑cell area of Jalapeño’s warp engine by roughly 13% and lifted performance‑power‑area (PPA) metrics on AMD’s GPUs by as much as 14%. In parallel, the AI‑driven workflow has been applied to RISC‑V core designs, further showcasing the breadth of the approach. The development matters because it creates a feedback loop in which AI tools that rely on high‑performance chips are now helping to build the next generation of those chips. If the reported productivity boost – described by OpenAI co‑founder Greg Brockman as compressing a year’s worth of design work into five months – holds up under independent verification, it could shorten development cycles, lower costs and accelerate the rollout of more capable AI hardware. That, in turn, may reshape the competitive dynamics among chipmakers and influence supply‑chain strategies that have recently been under pressure, such as Nvidia’s sales restrictions to China. The next step will be rigorous third‑party validation of the claimed area and performance improvements, as analysts caution against relying on specification‑based scores alone. Observers will also watch whether OpenAI extends the model‑driven workflow to other partners beyond AMD, and how traditional EDA vendors respond to a growing AI‑centric design paradigm.
102

Why Gemini 4 Carbon Is Delayed

Why Gemini 4 Carbon Is Delayed
HN +5 sources hn
benchmarksgemini
A private benchmark released this week shows Google’s Gemini 4 Carbon model struggling to keep pace with rival agents. In a series of bug‑fixing and code‑review tasks that competitors Opus and Astra completed in a few minutes, Gemini 4 Carbon required close to an hour, prompting users to label the service “nearly unusable.” Step‑level tracing performed by the benchmark’s authors pinpointed the bulk of the slowdown to two technical culprits. First, the model appears to over‑audit code that is unrelated to the immediate task, inflating inference time by 16‑30 % compared with Astra and 45‑49 % versus Opus. Second, mechanical inefficiencies such as fewer tool calls per step and frequent re‑reading of the same material further drag performance. The investigators estimate that addressing the tool‑call and re‑reading issues alone could shave roughly a quarter off the total runtime. The findings matter because Gemini 4 Carbon is positioned as Google’s flagship offering for more complex, multi‑step workflows. Persistent latency erodes the user experience that Google has been building with recent releases—such as the lower‑memory Android Gemini app and expanded Gemini Tasks—while giving competitors a clear performance edge. If the model remains sluggish, developers may gravitate toward faster alternatives, potentially slowing adoption of Google’s broader Gemini ecosystem. All eyes are now on Google’s response. The company has previously highlighted performance differences between its Gemini variants, and a forthcoming patch or architectural tweak could mitigate the identified bottlenecks. Monitoring upcoming updates, community‑driven workarounds and any official statements from Google will be key to gauging whether Gemini 4 Carbon can recover its competitiveness.
88

NYC, LA and Chicago school districts look at alternatives to Google Search after company refuses a version without AI features.

NYC, LA and Chicago school districts look at alternatives to Google Search after company refuses a version without AI features.
Techmeme +1 sources techmeme
google
NYC, Los Angeles and Chicago school districts are actively exploring search‑engine alternatives after Google refused to supply a version of its core Search service stripped of generative‑AI features. The districts, which serve millions of students, had asked for a “clean” search experience that would omit the chatbot‑style responses and other AI‑driven enhancements that Google has rolled out across its consumer platform. Google’s decision to keep those features bundled sparked a swift backlash from educators concerned about the impact on student learning, data privacy and the potential for misinformation. The move underscores a growing tension between the rapid rollout of AI‑powered tools and the need for safe, age‑appropriate digital resources in K‑12 environments. While Google argues that its AI adds value by summarising information and answering questions more naturally, school officials say the lack of an opt‑out option makes it difficult to align the service with district policies and parental expectations. What comes next will hinge on how quickly viable, privacy‑focused alternatives can be vetted and deployed. Districts are likely to pilot open‑source or niche commercial search engines that promise transparent data handling and the ability to disable AI layers. The outcome could set a precedent for other U.S. school systems grappling with similar demands for AI‑free tools. As the debate unfolds, it will also test whether major tech firms can accommodate the nuanced requirements of public education without compromising their broader AI strategies.
87

OpenAI reports agents now complete 46% of researchers' 2‑4‑hour tasks, up from 23% in March, with median daily agent spend at $2,400 (OpenAI)

OpenAI reports agents now complete 46% of researchers' 2‑4‑hour tasks, up from 23% in March, with median daily agent spend at $2,400 (OpenAI)
Techmeme +7 sources techmeme
agentsopenai
OpenAI has released a new internal study, “Research acceleration: the jagged frontier,” that quantifies how its coding agents are reshaping the pace of AI research. The data show that agents now finish 46 % of researchers’ tasks estimated to take two to four hours, up from 23 % in March. For longer assignments of four to eight hours, solo‑completion rates have risen from 13 % to 35 % over the same period. Success varies by task type – monitoring training runs is already 79 % automated, while more complex work lags behind. The report also reveals the financial side of the shift. The median researcher now spends roughly $2,400 per day on agent usage, reflecting both the growing reliance on these tools and the cost of the compute they consume. OpenAI stresses that, despite the gains, human oversight remains essential; agents are positioned as accelerators rather than replacements in the research loop. Why the numbers matter is twofold. First, they provide a concrete benchmark for how far AI‑assisted development has progressed, suggesting that a substantial share of routine, multi‑hour work can be delegated to software. Second, the cost signal hints at a new budgeting line for labs and companies that adopt such agents, potentially reshaping R&D spending patterns. Looking ahead, OpenAI is calling for mandatory disclosure of indicators when AI is used to develop other AI systems, a move that could set industry standards for transparency. Observers will watch whether the upward trend in agent success continues, how quickly the $2,400‑per‑day spend scales across the broader research community, and whether other firms – from chip designers to enterprise AI startups – adopt similar agent‑driven workflows. The next report, expected later this year, should reveal whether the current gains translate into faster publication cycles and more rapid model iteration.
64

Anthropic secures about $3 billion from Japanese investors in IPO, hitting the low end of its $3‑3.5 billion target

Anthropic secures about $3 billion from Japanese investors in IPO, hitting the low end of its $3‑3.5 billion target
Techmeme +6 sources techmeme
anthropic
Anthropic PBC has secured roughly $3 billion from Japanese investors in its U.S.‑linked initial public offering, a filing shows. The cash comes at the low end of the company’s target range of $3 billion to $3.5 billion and was raised by selling Class A shares at ¥83 ($0.60) each through six Japanese securities firms. The filing, submitted on Nov 2, confirms that the Japanese tranche covers only part of the broader Nasdaq listing, which still awaits regulatory clearance. The transaction underscores the appetite for AI‑focused equities in Japan, a market that recently backed SpaceX’s record‑size IPO with $2.2 billion. Anthropic’s ability to attract a full‑billion‑dollar slice of its offering suggests confidence in its Claude models and its positioning as a “safer” AI alternative. It also feeds speculation about the firm’s valuation, with some analysts floating a $1.5 trillion figure. The raise follows our Nov 18 report that Anthropic was considering upsizing its IPO price well above the $80‑per‑share range amid surging demand, and it arrives amid growing scrutiny from groups such as the Center for AI Safety, which have challenged the company’s safety claims. Investors will now watch for the Nasdaq filing’s approval, the final pricing of the remaining shares, and the total amount raised once the U.S. portion closes. Market reaction to the Japanese commitment could set the tone for the rest of the offering and signal how the broader AI sector will be valued in the coming months.
52

Gemini 4 Neon priced at $3 per million input tokens and $15 per million output tokens, positioned between Hydrogen and Carbon and replacing Argon in the Gemini app

Techmeme +1 sources techmeme
geminigoogle
Google has added a new tier to its Gemini 4 family, unveiling Gemini 4 Neon in the Gemini app. The model is priced at $3 per million input tokens and $15 per million output tokens, positioning it cost‑wise between the recently launched Hydrogen and Carbon variants. Neon also supersedes the earlier Argon model that powered the app’s conversational features. The rollout follows last week’s release of Gemini 4 Carbon and Hydrogen, which sparked discussion about Google’s tiered pricing strategy and the performance trade‑offs between the models. By inserting Neon between the two, Google gives developers and end‑users a mid‑range option that balances price and capability, potentially widening adoption for workloads that find Argon too limited but do not need the higher compute budget of Carbon. The pricing shift matters because token‑based charges are becoming a benchmark for AI service economics, influencing everything from startup budgeting to enterprise procurement. At $3/$15 per million tokens, Neon is markedly cheaper than Carbon’s rates (as reported in our earlier piece on Gemini 4 Carbon) while still promising more sophisticated output than Hydrogen. This could accelerate migration of existing Argon‑based integrations to the newer model, reshaping usage patterns within Google’s ecosystem. What to watch next: how developers respond to Neon’s performance and cost profile, and whether Google will adjust pricing or introduce additional tiers. Observers will also be keen to see if Neon’s launch prompts competitors to refine their own mid‑range offerings, and whether the model’s capabilities narrow the gap that has kept some users on alternative platforms. Further details on Neon’s technical specs and real‑world benchmarks are expected in the coming weeks.
51

Satya Nadella predicts AI agents will reshape web economics

TechCrunch +5 sources techcrunch
agentsamazonmicrosoft
Microsoft chief executive Satya Nadella told the New York Times Dealbook Summit that the rapid rise of AI‑driven agents will upend the economics of the internet. He warned that firms built on advertising‑driven “human web” traffic will have to move toward cost‑per‑click or transactional pricing models, likening the shift to the early‑2000s migration from portal sites such as Yahoo to search‑engine‑centric Google. Nadella singled out large platforms that rely on eyeball‑based ad revenue, noting that companies that block agents – for example, Amazon – risk losing market share to rivals that embrace agentic access. The implication is that AI agents could become the primary interface for users to locate products, services and information, bypassing traditional browsers and the ad impressions that currently fund much of the web’s ecosystem. The comment arrives as AI agents are already proving their commercial value. OpenAI reported that its agents now complete nearly half of researchers’ routine tasks, while KPMG’s pilot using more than a hundred agents to audit a FTSE 250 firm demonstrated deeper questioning than human staff. Those developments suggest the technology is moving from experimental to revenue‑generating roles, making Nadella’s warning timely for advertisers and platform operators alike. What to watch next are the strategic responses from ad‑centric firms. Will Amazon and others open their sites to agents, or double down on blocking? Industry observers will also monitor whether Microsoft rolls out new tools to help developers integrate agents, and how regulators react to a web economy that could shift from impression‑based to transaction‑based monetisation. As we reported on the growing deployment of AI agents in research and enterprise [2026‑11‑19], the next few months could define the new commercial architecture of the internet.
48

Managing frontier AI risk from training to deployment

OpenAI +6 sources openai
alignmentopenaitraining
OpenAI has published a detailed account of how it reins in the risks posed by its most powerful models, outlining a lifecycle that stretches from early‑stage training to public deployment. The company’s “Managing frontier AI risk: From training to deployment” brief, released on 18 November 2026, maps a sequence of controls – model alignment, sandboxed research environments, blocking safeguards and continuous monitoring – and spells out explicit pause points whenever alignment evidence or containment falls short. The narrative also describes how frontier models transition from tightly‑controlled internal experiments to employee‑facing ChatGPT and Codex tests before reaching public APIs. The disclosure matters because it marks the most comprehensive internal risk‑management roadmap OpenAI has shared since the Hugging Face incident forced the industry to rethink containment practices. By making its safeguards public, OpenAI signals a willingness to be transparent about the safety levers it relies on, a move that could shape regulatory expectations and industry standards for frontier AI governance. The timing also dovetails with the broader debate sparked by recent coverage of the frontier‑AI pause, where chip makers and other stakeholders have voiced skepticism about voluntary halts. As we reported on 18 November 2026, “Remember the frontier AI pause? Chip companies don’t either,” highlighting the pressure on AI developers to demonstrate robust self‑regulation. What to watch next are the practical outcomes of OpenAI’s framework. Regulators may cite the document when drafting oversight rules, while competitors could adopt similar sandboxing and pause protocols to pre‑empt scrutiny. Observers will also track whether the stated safeguards affect rollout speed for new features, and whether any future alignment failures trigger the explicit pauses OpenAI has built into its pipeline. The evolution of these controls will be a key barometer for the industry’s ability to balance rapid innovation with safety.
48

OpenAI's AI discovers dozens of flaws in Apple iPhones Macs

Mastodon +6 sources mastodon
appleopenai
OpenAI’s newest defensive‑oriented model, GPT‑6 Cyber, has been put to work on Apple’s flagship devices. In a collaboration with cybersecurity firm Calif, researchers ran the model over recent months and uncovered roughly 35 security flaws spanning iPhone and Mac hardware and software, according to a Forbes report published on 17 November. Some of the vulnerabilities could let an attacker seize control of a device simply by luring a user to a malicious website. The discovery underscores how AI is reshaping the vulnerability‑research landscape. By automating parts of the hunt, GPT‑6 Cyber can scan large codebases and hardware specifications far faster than traditional manual audits, surfacing issues that might otherwise linger undetected. For Apple, the findings arrive at a moment when the company is tightening its own AI‑related safeguards – just a day earlier we reported on new disk‑control measures designed to keep AI agents from running unchecked on its platforms. What follows will be closely watched. Apple has not yet detailed a remediation timeline, but the company typically rolls out emergency patches through its regular software updates. Observers will be looking for a coordinated disclosure process, the scope of patches required, and whether Apple will adopt AI‑driven testing internally to stay ahead of similar threats. The episode also raises broader questions about the role of large‑scale language models in both exposing and potentially exploiting security gaps, a debate likely to surface in upcoming policy discussions and industry forums.
45

Gemini 4 Neon appears to rival Opus 5.5

HN +5 sources hn
agentsbenchmarksclaudegemini
Google has rolled out Gemini 4 Neon, the latest addition to its Gemini 4 family, and early benchmark data suggest the model is poised to challenge Anthropic’s Claude Opus 5.5 as the leading high‑performance LLM. Based on the performance of Gemini 4 Carbon – a reasonable proxy given Neon’s even better scores – analysts project Neon would achieve an Elo rating of roughly 1,065 on the Agent Duel leaderboard, just shy of Opus 5.5’s 1,078 in the same snapshot. Google’s own benchmark table, released alongside the model, shows Neon ahead of both Opus 5.5 and GPT‑6.2 Sol on the majority of tests, leading on 16 of 19 evaluations, including the Artificial Analysis suite. The comparison matters because Opus 5.5 has become a reference point for enterprise‑grade reasoning and coding tasks. If Neon can consistently out‑perform it across a broad set of metrics, developers and businesses may gravitate toward Google’s ecosystem for cost‑effective, high‑throughput workloads. The model’s projected 122 tokens‑per‑second throughput further underscores its suitability for real‑time applications. As we reported on 19 November, Neon is priced at $3 per million input tokens and $15 per million output tokens, sitting between Hydrogen and Carbon and replacing Argon in the Gemini app. The next steps will be to see how the projected Elo translates into real‑world usage, whether latency and context‑window claims hold up under production loads, and how Google’s pricing strategy competes with Anthropic’s offerings. Follow‑up testing from independent labs and early adopter feedback will be the clearest barometer of Neon’s market impact.
32

VideoNexus: Video Understanding Moves Beyond Reasoning Context

HF Papers +5 sources hf papers
agentsreasoningtraining
A new framework called **VideoNexus** has been unveiled, promising a more efficient way for AI systems to understand long, dense video sequences. The authors – Yancheng Long, Xuanyu Zheng, Wei Chen and colleagues – present VideoNexus as a **training‑free** architecture that cleanly separates two functions that have traditionally been tangled: fine‑grained video perception and evidence‑driven planning and reasoning. The core idea is to let a “planner” decide **where** to look, **what** visual details to resolve, and **how densely** to sample frames, while a separate perception sub‑agent handles the heavy lifting of extracting those details. By exchanging only compact evidence rather than feeding the entire visual stream into a monolithic model, the system can focus computational effort on the most relevant parts of a video. The authors argue that this approach yields better long‑video reasoning than the common practice of indiscriminate context accumulation. Why it matters: interpreting extended video content remains a bottleneck for AI, as the dual demands of perception and reasoning quickly overwhelm current models. VideoNexus offers a modular alternative that could reduce latency, lower memory footprints, and improve answer fidelity when answering questions that require locating and integrating visual evidence across time. The concept dovetails with recent advances in video agents and interactive‑latency generation that we covered in our FastFrames and Beyond Answering pieces earlier this week. What to watch next: the research team has released implementation resources and reproducibility signals, inviting the community to test the planner‑perception split on benchmark tasks. Follow‑up work may explore integration with existing video‑question‑answering pipelines, scaling the planner’s summarisation capabilities, and open‑source releases that could accelerate adoption across Nordic AI labs and beyond.
30

AI Language Models Can Self‑Edit Their Weights

ArXiv +1 sources arxiv
inferencetraining
A new pre‑print on arXiv (2611.21885v1) claims that a language model can rewrite its own weights while running, breaking with the long‑standing practice of training a model once, freezing it, and then deploying it unchanged. The authors demonstrate a “frontier” model that, given a task, modifies the parameters that underlie its own computations rather than relying on an external prompt, scaffold or extra inference‑time processing that disappears once the request is fulfilled. Why it matters is twofold. First, self‑editing weights could let a single model continuously improve or specialise without the costly, time‑intensive retraining cycles that dominate current AI development. Second, the ability to adapt internal representations on the fly opens a path toward more autonomous agents that can learn from experience in situ, a capability hinted at in OpenAI’s recent report on agents completing a growing share of researchers’ routine tasks. If models can safely and efficiently update themselves, the economics of AI deployment could shift dramatically, lowering compute budgets and shortening time‑to‑market for niche applications. What to watch next are the practical and safety dimensions of this approach. Researchers will need to prove that self‑modifying models remain stable, avoid catastrophic forgetting, and do not introduce hidden backdoors. Industry players may begin experimenting with such models in low‑risk settings, while regulators and ethicists will likely scrutinise the implications of systems that can alter their own codebase. Follow‑up studies and open‑source implementations will determine whether weight‑editing becomes a mainstream tool or remains a laboratory curiosity.
28

Google launches Gemini 4 Neon, promising fast end‑to‑end task completion for coding, knowledge work and science in API, Antigravity, and the Gemini app

Techmeme +6 sources techmeme
deepmindgeminigoogle
Google has added a new tier to its Gemini family – Gemini 4 Neon – and made it available today through the Gemini API, the Antigravity interface and the Gemini app. The model is marketed as delivering “fast end‑to‑end task completion” for complex software engineering, knowledge‑work and scientific workflows. In Google’s own launch notes, Neon is positioned between the recently released Gemini 4 Carbon (aimed at long‑horizon projects) and Gemini 4 Hydrogen (optimised for rapid, everyday tasks), offering a blend of high‑performance coding, multi‑step reasoning and low‑latency agentic execution. The rollout matters because it pushes Google’s generative‑AI offering further up the performance curve, targeting developers and enterprises that need more than simple text generation. By embedding Neon in Antigravity and the Gemini app, Google is extending its AI capabilities from cloud‑only APIs to interactive, on‑device experiences, potentially accelerating adoption in productivity tools and research environments. The move also tightens competition with other frontier models; as we reported on 19 Nov 2026, Neon appears to be positioned as a rival to the Opus 5.5 class of models. What to watch next is how Google prices Neon and whether it will match the $3‑$15 per‑million‑token rates disclosed for Hydrogen and Carbon. Early performance benchmarks and real‑world case studies will reveal whether Neon can deliver the promised speed gains in multi‑step agent workflows. Integration with the Gemini Enterprise Agent Platform will be a key test of its appeal to large organisations, while developers will be looking for SDK updates and tooling support. Follow‑up reporting will track adoption rates, pricing announcements and any competitive responses from other AI providers.
27

Google introduces incognito mode for AI search feature

TechCrunch +5 sources techcrunch
google
Google has rolled out an “incognito‑style” option for the AI‑powered conversations that sit inside Search. The new feature, called temporary threads, lets users start a dialogue in AI Mode that is not recorded in their search history and does not feed into future recommendation algorithms. The change is live globally on both desktop and mobile versions of Search. The move addresses growing concerns over how conversational data are harvested for personalization and advertising. By keeping a chat isolated, Google gives users a way to ask sensitive or one‑off questions without leaving a permanent trace in their account. The design mirrors a similar privacy tool already used in the Gemini suite, signalling that Google is extending its “private‑by‑default” thinking across its AI products. The addition could reshape how people interact with Search’s generative layer. Users who were hesitant to experiment with AI Mode because of data‑retention worries may now feel freer to explore, potentially boosting engagement with the feature. At the same time, the temporary‑thread model may limit the system’s ability to learn from individual interactions, which could affect the relevance of personalized results and the value of ad targeting that relies on search histories. What to watch next includes how Google defines the lifespan of these temporary threads, whether users can export or delete them manually, and if the company will expand the privacy controls to other AI services such as Gemini. Regulators and privacy advocates will likely scrutinize the implementation, while competitors may roll out comparable safeguards to keep pace in the rapidly evolving search‑AI landscape.
27

Palmer Luckey warns AI‑powered cyberattacks outpace defenses, urges action

TechCrunch +1 sources techcrunch
Palmer Luckey, the Anduril co‑founder, has warned that AI‑powered cyberattacks are already outpacing existing defences. Speaking at a recent industry forum, Luckey said the reliance on legacy hardware and software is leaving critical systems exposed to a new wave of attacks that can be generated, refined and deployed by artificial‑intelligence tools at unprecedented speed. “It’s time to start running,” he warned, urging organisations to treat AI‑enabled threats as a strategic priority rather than a technical footnote. The comment adds urgency to concerns raised only days earlier when Microsoft’s Satya Nadella warned that large swathes of critical infrastructure are effectively “unpatchable” in the face of sophisticated AI exploits. Both remarks underscore a growing consensus that traditional security models—largely built around manual patch cycles and signature‑based detection—are ill‑suited to adversaries that can automate vulnerability discovery, craft polymorphic malware and adapt tactics in real time. Why the alarm matters now is twofold. First, the democratisation of generative AI lowers the barrier for less‑skilled actors to launch potent attacks, expanding the threat landscape beyond nation‑state actors. Second, many enterprises and public‑sector networks still run outdated operating systems and network devices that cannot be upgraded without costly overhauls, creating a fertile ground for AI‑driven exploitation. What to watch next are the concrete steps the defence and technology sectors will take. Industry analysts expect a surge in investment in AI‑augmented security platforms, while regulators may accelerate proposals for mandatory AI‑risk assessments in critical sectors. Anduril itself is likely to showcase its own AI‑defence solutions, positioning the company at the forefront of a market that is rapidly shifting from reactive patching to proactive, intelligence‑driven protection.
27

Satya Nadella warns critical infrastructure remains unpatchable amid fears of AI cyberattacks

TechCrunch +5 sources techcrunch
microsoft
Microsoft’s chief executive used the New York Times DealBook Summit on Wednesday to warn that a large share of the United States’ critical‑infrastructure software is effectively “unpatchable.” Satya Nadella said that, in an era where artificial‑intelligence tools can locate flaws at machine speed, many legacy systems that run power grids, transportation networks and building‑automation equipment cannot be updated fast enough to keep pace with emerging threats. He cited the recent Outlook outage as a reminder that even cloud‑centric services can be disrupted, but emphasized that AI also creates new defensive possibilities alongside the risks. The comment matters because AI‑driven bug hunting is already reshaping the vulnerability‑management landscape. Earlier this month Anthropic trimmed its coordinated‑disclosure window from 90 to 30 days after AI‑assisted researchers began surfacing high‑severity bugs more quickly, a shift that underscores how quickly attackers could exploit the same capabilities. If core infrastructure cannot receive timely patches, a wave of AI‑generated exploits could trigger a “Y2K‑style” cascade of failures in everything from HVAC chillers to vehicle control units. What to watch next: Microsoft is likely to double down on AI‑powered security tools, a theme echoed in its recent rollout of Word, Excel and PowerPoint inside Copilot, which the CEO has called “the new OS for work.” Regulators may also step in, as policymakers grapple with the prospect of widespread, unmitigated vulnerabilities. Industry observers will be tracking whether Microsoft announces new remediation frameworks, firmware‑update standards or collaborations with governments to harden the most stubborn legacy systems.
25

π^4 Foundation Model Advances 4D Scene Geometry and Motion

HF Papers +5 sources hf papers
A research team spanning Shanghai AI Laboratory, Shanghai Jiao‑Tong University and Zhejiang University has unveiled π⁴, a unified foundation model that simultaneously reconstructs 3‑D geometry and estimates motion from raw RGB inputs. The system accepts a single photograph, an unordered set of images or a video clip, and in a single forward pass produces dense scene geometry, metric‑scaled camera poses, 2‑D/3‑D point tracks and masks that isolate moving objects. The model’s architecture pairs a geometry branch with a geometry‑conditioned motion branch, allowing the two tasks to inform each other. At roughly 0.5 billion parameters, π⁴ is compact enough to run as a single model while delivering strong performance across a range of perception tasks, according to the authors’ pre‑print and accompanying GitHub repository. All code, a project page and a Hugging Face demo have been released publicly. π⁴ matters because it collapses what has traditionally been a multi‑stage pipeline—calibrated multi‑view capture, separate structure‑from‑motion, and dedicated tracking—into one end‑to‑end system that works on uncalibrated, “in‑the‑wild” footage. This lowers the barrier for applications such as augmented reality, robotics navigation and large‑scale mapping, where acquiring calibrated image streams is often impractical. The model also exemplifies the broader shift toward foundation models that can be fine‑tuned for diverse visual tasks, echoing recent advances in self‑modifying language models and cost‑effective large‑scale vision services. The community will now watch for benchmark results that compare π⁴ against existing 3‑D reconstruction and motion‑estimation suites, as well as for integration into downstream products. Early adoption by AR/VR developers or robotics platforms could signal the model’s practical impact, while further scaling or specialization may extend its reach into more demanding industrial scenarios.
24

NEBench R1: New Benchmark Tests Nuclear Expertise in Large Language Models

ArXiv +5 sources arxiv
ai-safetybenchmarksreasoning
A new benchmark called NEBench R1 has been released to test how well large language models (LLMs) can reason through nuclear‑engineering problems. The dataset pulls together 7,546 distinct multiple‑choice questions taken from U.S. Nuclear Regulatory Commission (NRC) reactor‑operator licensing exams, giving the benchmark a breadth that surpasses earlier nuclear‑domain tests, which tended to cover only narrow topics or focus solely on whether a model’s final answer was correct. The study reports that six leading LLMs achieved accuracy scores ranging from 34.0 % to 73.9 % on the NRC questions. More importantly, NEBench R1 is designed to assess the quality of the models’ underlying technical justification, not just the end result. This focus on reasoning is intended to address safety‑critical concerns: in nuclear engineering, a correct answer without a sound explanatory chain can be as risky as a wrong one. The benchmark arrives as the industry explores LLMs for knowledge‑intensive, high‑stakes applications, from design assistance to operational decision support. By providing a systematic way to gauge both correctness and logical rigor, NEBench R1 offers regulators, plant operators and AI developers a clearer picture of where current models stand and where they fall short. Going forward, researchers are likely to apply the benchmark to newer model versions, expand it with additional exam sections, and explore how performance correlates with real‑world safety outcomes. Watch for follow‑up studies that compare model improvements, as well as any uptake of NEBench R1 by nuclear authorities as a reference point for AI‑assisted tooling in the sector.
16

Meta pauses Class C share issuance amid investor lawsuits over funding AI investments without diluting Zuckerberg's stake

Techmeme +1 sources techmeme
meta
Meta Platforms Inc. announced that it will temporarily suspend the issuance of a new non‑voting Class C share class. The pause comes as a group of investors has filed lawsuits challenging Meta’s plan to raise capital for its expanding artificial‑intelligence portfolio without diluting Mark Zuckerberg’s ownership stake. The disputed share class would have allowed Meta to tap public markets for AI funding while keeping voting power concentrated in the founder’s hands. Investors argue that the structure undermines shareholder rights and could set a precedent for other tech firms seeking to separate economic interest from control. By halting the rollout, Meta signals a willingness to address the legal challenge while it evaluates alternative financing routes. The move matters because Meta’s AI ambitions are a central pillar of its next‑generation growth strategy, and the company’s ability to fund them without a major equity dilution could influence how other high‑profile tech firms structure future capital raises. The pause also highlights growing scrutiny of dual‑class share structures, especially when they are used to finance fast‑moving, capital‑intensive projects such as large‑scale AI development. Going forward, observers will watch how the lawsuits progress and whether Meta proposes a revised share‑issuance plan that satisfies both investors and regulators. The outcome could shape the broader debate over corporate governance in the AI era and affect Meta’s timeline for rolling out new AI products and services.
15

Trump turns new Philippine land deal into AI hub

The Verge +1 sources the verge
apple
President Donald Trump touched down in the Philippines on Tuesday to a flamboyant welcome that included five beauty queens, a child waving both U.S. and Philippine flags, and a marching band playing “YMCA,” a staple of his rally playlists, according to local outlet Rappler. The ceremony, staged on the tarmac, underscored the political theater surrounding his latest overseas venture, which the headline frames as a “land grab” intended to become an AI hub. The announcement signals a new chapter in Trump’s post‑presidential business pursuits, this time targeting the fast‑growing artificial‑intelligence sector in Southeast Asia. While the snippet offers no specifics on the size of the parcel, the technology focus, or any partnership arrangements, the implication is that the site could serve as a base for AI development, data processing or related services. If realized, the project would add a high‑tech dimension to the traditionally infrastructure‑focused foreign investments in the region and could deepen U.S. tech ties with the Philippines. Why it matters is twofold: first, it places a high‑profile American figure at the centre of a strategic technology push in a country that is already courting AI talent and investment; second, it raises questions about regulatory oversight, data sovereignty and the geopolitical optics of a former U.S. president spearheading a private AI enclave abroad. Observers will be watching for official statements from the Trump organization, details on the land transaction, any involvement of local or U.S. government agencies, and the reaction of Philippine tech firms and regulators. Follow‑up reporting will focus on the project's concrete plans, financing and timeline, as well as its potential impact on the regional AI ecosystem.

All dates