AI, summarized

Claude-assisted Riemann proof claimed, Anthropic closes record IPO, open models multiply

An Anthropic mathematician says Claude helped produce a Lean-formalized proof of the Riemann Hypothesis, reportedly confirmed by the Lean community, though confirmation rests on secondary reports. Anthropic also closed a roughly $93.5 billion IPO. Elsewhere, a canary study exposed leaky AI routers, and Thinking Machines' looped Inkling 2 Mini arrived with immediate tooling support.

38 stories, chosen from 90 new entries. Written 5 minutes ago.

Ranked stories

The heavier the headline, the higher the impact:

Show a category
  1. 1 Major

    Anthropic mathematician claims Claude-assisted Riemann hypothesis proof; Lean community reportedly confirms it, rival human draft appears

    Levent Alpöge of Anthropic announced a proof of the Riemann Hypothesis, and of the Generalized version, found working with Claude and released as Lean code of roughly 250,000 page equivalents. Hours later Stanford circulated an all-human draft by Alessandro Masullo, which others flagged problems in. The Decoder reports the Lean and Mathlib community has confirmed the Lean proof.

    Why it matters. If it holds, an AI system has solved the most famous open problem in mathematics, a major capability milestone. Confirmation so far rests on secondary reports, and the formalized statement still needs checking against the original problem.

    ResearchImpact88Fit82Score 93

    Reported by The Verge, TechCrunch, The Decoder, Simon Willison and Latent Space

    A Claude-driven, Lean-verified solution to the Riemann hypothesis would be a landmark AI capability result, though confirmation comes via reports; highly relevant to a reader following frontier models.

  2. 2

    Thinking Machines releases Inkling 2 Mini, an open looped transformer that runs on a 32GB Mac

    Thinking Machines released Inkling 2 Mini under Apache 2.0. It is a 26.4B-parameter mixture-of-experts model that runs 12 of its 24 layers three times per token, giving 5.4B effective active parameters and a 1M-token context. It scores 20 on Artificial Analysis's Intelligence Index, tying Qwen3.8 27B, and reports 48.2% on DeepSWE v1.1.

    Why it matters. A looped design that reasons internally without extra tokens could give developers a cheap local model for agentic coding. The benchmark claims are mostly vendor-reported, apart from the Artificial Analysis score.

    ModelsImpact62Fit90Score 77

    Reported by The Decoder, MarkTechPost and Hacker News (478 points)

    A new open-weight model with an architecture twist, strong small-model agentic coding results and local deployment is highly relevant to builders. Impact is held back by mid-tier absolute scores and vendor-sourced claims.

  3. 3

    Paper claims open 27B model improves itself by rewriting its own weights without gradients

    Researchers led by Jiajun Fan report that an open 27B model, given its weights, graded answers and the ability to run itself, edited its own weights without gradients or training. They say accuracy rose 6.6 points on unseen problems, kept rising over eight sessions, and the model repaired a damaged copy to over 80% of original accuracy on image tasks.

    Why it matters. If it holds up, weight self-editing at inference-style compute would be a step toward recursive self-improvement, with safety implications. It is a single unreviewed arXiv preprint, so the claims need independent replication.

    ResearchImpact58Fit80Score 69

    Reported by Hacker News (487 points)

    Novel but unreplicated preprint claiming gradient-free self-rewriting and a path to recursive self-improvement; closely tied to model capabilities and AI risk, which the reader follows closely.

  4. 4

    Researcher's 2 million canary prompts reveal credential theft and privacy failures across AI routers and resellers

    HD Moore of Dragos sent nearly two million prompts with unique trackable secrets through AI routers and hosting providers. Secrets leaked: at least three providers were compromised, some resellers logged and reused credentials, and a Chinese lab scraped URLs despite zero-data-retention flags. OpenRouter later removed a ZDR label from MiniMax endpoints.

    Why it matters. Developers using AI routers or reseller endpoints cannot assume privacy flags like Zero Data Retention are honored. Secrets in prompts may be stolen or reused, so choosing upstream providers directly matters.

    SafetyImpact62Fit66Score 64

    Reported by Ars Technica and Lobsters

    Large, concrete experiment showing real credential theft and broken ZDR promises in AI inference supply chain; directly relevant to developers using routers and APIs, though not model or alignment work.

  5. 5

    Google previews Tool Retrieval in Gemini API to defer tool schemas for large agent tool libraries

    Google launched Tool Retrieval in preview in the Gemini Interactions API (v1beta). Agents mark functions or MCP servers with defer_loading, and the model sees only names and short descriptions until it requests a schema. Server-side mode resolves search, load and call in one request; client-side mode lets apps use their own registry at one extra round trip.

    Why it matters. Large tool libraries and MCP setups can burn tens of thousands of tokens per request, and Gemini now matches Anthropic and OpenAI tool search. Google published no savings data, so real gains are unproven.

    ToolsImpact48Fit88Score 64

    Reported by MarkTechPost

    A meaningful agent-tooling API feature that follows Anthropic and OpenAI, with no published savings and still in preview; it squarely matches the reader's developer-tools and API interests.

  6. 6

    Ollama v0.40.17 adds Thinking Machines' Inkling 2 Mini, supported on Apple silicon first

    Ollama's v0.40.17 release adds Inkling 2 Mini from Thinking Machines, a 26B-parameter model with 5.4B active parameters and looped layers. It needs about 20GB, handles text and images, and scores 79.2% on SWE-Bench Verified. It runs only on Apple silicon for now, and the release also adds Claude Code deferred tool loading and concurrent requests on Apple silicon.

    Why it matters. A small model with a reported 79.2% on SWE-Bench Verified that runs locally on a Mac could suit local coding agents. The score comes from release notes and is not independently verified.

    ModelsImpact52Fit80Score 63

    Reported by Ollama releases

    A new locally runnable model with a strong claimed coding benchmark score plus agent tooling updates is relevant to builders, though the score is unverified and PC support is missing.

  7. 7

    NVIDIA releases Nemotron-Labs-Sigma, open continuous diffusion language models at 3B and 8B

    NVIDIA published 14 open checkpoints of Nemotron-Labs-Sigma, diffusion language models that denoise text in a 16-dimensional continuous embedding space instead of over discrete tokens. Sigma-8B base averages 71.16 on NVIDIA's six-task suite, above Qwen3-8B's 68.91. A self-speculative 8B variant decodes 3.46x faster than autoregressive decoding on a GB200.

    Why it matters. It is a large-scale open test of continuous diffusion for text, with distilled and speculative variants that cut inference steps. All benchmarks are NVIDIA's own, and hard reasoning, context length and GPU-only support remain limits.

    ModelsImpact55Fit72Score 62

    Reported by MarkTechPost

    A notable open model family from a major lab with a novel architecture and speedups, though self-reported benchmarks and small scale limit impact; it matches the reader's top interest in releases.

  8. 8

    METR updates AI time horizon plot with longer, more open-ended judgment tasks

    METR released an updated time horizon report that adds tasks estimated to take humans far longer and tasks that are more open-ended. Capabilities on verifiable tasks keep improving exponentially, but models fall short on judgment-heavy work such as designing useful AI evaluations. Newer models like Opus 5.5 reach "borderline above the bar" on that task.

    Why it matters. It gives evidence on whether progress in coding carries over to fuzzier research skills, which bears on how soon AI R&D could be automated. Human completion times are mostly estimated and few such tasks exist, so the conclusions are tentative.

    ResearchImpact52Fit78Score 62

    Reported by MIT Technology Review

    Solid evaluation update with new information on judgment tasks and automated AI R&D, relevant to capabilities and risk, though limited by few tasks and estimated times.

  9. 9

    Transformers v5.24.0 adds Inkling 2 Mini, Apertus 1.5, Dory and several other models

    Hugging Face released Transformers v5.24.0 with new architectures: Thinking Machines Lab's Inkling 2 Mini (looped layers), Swiss AI's Apertus 1.5, NVIDIA's Dory, Meta's Omnilingual ASR, Sarvam's 105B and 30B models, and EXAONE. It also renames past_key_values to cache across the library, with deprecation warnings and a peft 0.22.0 requirement.

    Why it matters. Developers get library support for many new open models in one release. The cache renaming is a breaking change that custom model code and Cache subclasses must adapt to eventually.

    ToolsImpact52Fit72Score 60

    Reported by Transformers releases

    A library release adding many open models and a notable API rename is useful to builders, but it is incremental, with no major new capability itself.

  10. 10

    Bytedance and US universities release Axion 1.0, an open agent model trained in three phases

    Bytedance and several US universities introduced Axion 1.0, an AI agent model trained in only three phases. They report that it rivals top commercial models from OpenAI and Anthropic on benchmarks. They credit carefully curated training data, and the model, data and code are freely available.

    Why it matters. If the results hold up, simpler and cheaper training pipelines could make competitive agent models easier to reproduce. Benchmark claims come from the authors and still need independent verification.

    ModelsImpact50Fit72Score 59

    Reported by The Decoder

    An open agent model with released data and code that claims parity with commercial models is notable and relevant to builders, though the benchmark claims are not independently verified.

  11. 11

    RadixArk releases DaisyRose-27B, an RL-only open-weight agentic coding model built on Qwen3.8-27B

    RadixArk released its first open-weight model, DaisyRose-27B, under Apache 2.0. It was post-trained from Qwen3.8-27B with reinforcement learning only, using about 1,200 tasks and 144 H200 GPUs over 18 days. Self-reported Terminal Bench 4.1 pass@1 rose from 5.6 to 12.0, with smaller gains elsewhere.

    Why it matters. A single-GPU open model with RL-driven agentic coding gains, plus an open RL framework (Miles), is useful for builders. All benchmark scores are self-reported, and it still trails much larger models.

    ModelsImpact45Fit80Score 59

    Reported by MarkTechPost

    An open-weight agentic coding model with RL-only training and deployment details fits the reader's top interests, but gains are modest, self-reported, and it trails larger models.

  12. 12

    OpenAI releases GPT-Live-1.1 voice model in API with cheaper idle billing

    OpenAI made GPT-Live-1.1 available in its API as gpt-live-1.1. It is a full-duplex multilingual voice model that can listen while speaking. OpenAI says accents are more natural and consistent across languages and tasks complete more reliably. A new standby billing state charges about $0.02 per minute for idle sessions.

    Why it matters. Developers running voice agents get a new model and lower costs for idle call time. Claims about accent quality and reliability come from OpenAI and are not independently verified in this report.

    ModelsImpact48Fit72Score 58

    Reported by MarkTechPost

    A new API voice model with pricing changes is a notable release for builders, but it is an incremental version update and the claims are vendor-reported.

  13. 13

    Anthropic closes IPO raising about $93.5 billion, plus $26 billion private placement

    Anthropic closed its IPO of 1.127 billion Class A shares at $83 each, including the underwriters' full option, for roughly $93.5 billion gross; about $91.6 billion went to Anthropic. A concurrent $26.0 billion private placement of Class N shares went to AMD, Broadcom, Google, Micron, Microsoft, NVIDIA and TSMC. Trading began on Nasdaq as "ANT" on November 19, 2026.

    Why it matters. A record-scale raise gives Anthropic large capital for compute and frontier model development, and ties key chip and cloud suppliers to it as shareholders. Effects on its products and safety posture are not yet clear.

    BusinessImpact66Fit38Score 57

    Reported by Anthropic and Hacker News (513 points)

    A very large IPO and supplier investments reshape industry finances, but it is a financing event; the reader's interests rank funding and market news low, with no model or safety content.

  14. 14

    OpenComputer launches as open-source, self-hostable Slack-based coding agent alternative to Devin

    OpenComputer is an Apache 2.0 coding agent that teams trigger by mentioning @Computer in Slack. It runs Claude Code, Codex or OpenCode in rootless Podman sessions and delivers pull requests. Its own benchmark claims PRs open 3.3x faster than Devin (56s vs 185s).

    Why it matters. It gives teams a self-hostable option for cloud coding agents, though the speed claims come from the vendor's own benchmark and are unverified.

    ToolsImpact40Fit78Score 56

    Reported by Hacker News (120 points)

    A useful open-source coding agent with self-hosting and multiple backends, but speed claims are self-reported and it is a small launch; it matches the reader's developer-tools interest closely.

  15. 15

    Google launches Duplex in Gemini Tasks, letting its agent phone businesses for users

    Google Research described how Duplex, introduced nearly a decade ago, moved from closed-domain recurrent models to hybrid LaMDA stacks and now a unified Gemini pipeline. This week it launched Duplex in Tasks in the Gemini app, which can call businesses on a user's behalf. Calls disclose they come from an AI, and merchants can opt out.

    Why it matters. Phone calling gives agents a fallback where no API or booking form exists. It shows how a production voice agent handles latency and guardrails, though the article is Google's own account and offers no independent evaluation.

    ToolsImpact50Fit62Score 55

    Reported by Google Research

    A meaningful agent product launch with useful architecture detail on voice agents, but largely a vendor retrospective without benchmarks, so mid impact and moderately relevant.

  16. 16

    California proposes AI emergency shutdown powers as federal RSI bills advance

    Transformer's weekly briefing reports a California commissioned report urging state authority to cap recursive self-improvement and order emergency shutdowns, plus a revised federal FRONTIER ACT from Rep. Trahan and a teased SAFE AI Act. It also notes new US licensing rules for overseas data centers serving Chinese firms.

    Why it matters. Concrete state and federal proposals aimed at recursive self-improvement could shape frontier lab obligations. The briefing mixes reporting with opinion, and claims about imminent RSI and Anthropic's IPO are unverified here.

    PolicyImpact48Fit62Score 54

    Reported by Transformer

    Concrete AI safety governance proposals matter for frontier model regulation, but they are proposals and the source is a newsletter roundup with commentary, so impact is moderate.

  17. 17

    Artificial Analysis: OpenAI's GPT-6.2 models give shorter answers but admit uncertainty less than Claude

    Artificial Analysis's AA-PersonalityBench finds OpenAI's GPT-6.2 Astra and Sol cut average answer length by 40 and 34 percent, to 210 and 223 words. Their Admits Uncertainty scores fell from 96 to 24 and from 98 to 51. Anthropic's Claude Fable 5.2 and Opus 5.5 score higher on Transparency.

    Why it matters. Style shifts in OpenAI's newest models affect how reliable and verbose answers feel in products built on them. The scores rate style relative to other models, not accuracy, so real-world error rates remain unclear.

    ModelsImpact42Fit72Score 54

    Reported by The Decoder

    A third-party benchmark of behavior changes in new OpenAI models versus Claude is relevant to model-focused builders, but it measures style rather than capability or accuracy.

  18. 18

    Columbia and CertiK release TensorLean to formally verify GPU kernels, finding bugs in 149 tested instances

    Researchers at Columbia University and CertiK introduced TensorLean, a Lean-based verifier that checks compiled PTX GPU kernels against tensor-level math specs. They tested 149 kernel instances, including FlashAttention-2, -3 and -4 and DeepGEMM. 83 were proved correct, and the other 66 turned up 38 correctness bugs. Specs run about 10,000 lines against 151,000 lines of kernel code.

    Why it matters. As more GPU kernel code is AI-generated, machine-checked proofs of correctness could catch subtle bugs in widely used attention and GEMM kernels. Practical adoption depends on spec-writing effort and coverage beyond these kernels.

    ResearchImpact48Fit58Score 53

    Reported by MarkTechPost and Latent Space

    Solid research that found real bugs in widely used kernels, relevant to developers building with AI, but not a model release or alignment work.

  19. 19

    OpenAI cuts reasoning-extraction attack success on GPT-6.2 models from 54% to 0.2%

    OpenAI strengthened safeguards against extraction of protected internal reasoning from GPT-6.2 Astra and GPT-6.2 Sol. In its adaptive red-teaming, extraction from prior requests fell from 54.0% to 0.2%, and from active requests from 22.9% to 0.9%. It also bans thousands to tens of thousands of accounts weekly over distillation campaigns.

    Why it matters. Reasoning traces are a main target for distillation, so tighter protection limits what developers and rivals can see or copy. The figures come from OpenAI's own tests and are not independently verified.

    SafetyImpact45Fit62Score 52

    Reported by The Decoder

    A notable misuse-defense update with concrete vendor-reported numbers on frontier models, relevant to API builders and model security, but self-reported and not a new capability.

  20. 20

    Nvidia researchers argue summaries and search can replace very large context windows

    Nvidia researchers say language models often do not need their full context windows, which are slow, costly and error-prone. In their tests, periodic summaries and targeted keyword lookups usually matched full-context results at much lower cost. They also propose a model class that works from short notes rather than long context.

    Why it matters. If the results hold, developers could cut inference cost and latency by using summarization and retrieval instead of long prompts. The findings come from the authors' own tests and are not independently verified.

    ResearchImpact42Fit66Score 52

    Reported by The Decoder

    Solid research with practical cost implications for building with LLMs, but it is a single-source report on unreplicated tests, so impact is moderate; fit is good for model and tooling interests.

  21. 21

    Baseten releases open-weight decision models bison-1 and groundhog-1 built on Gemma 4

    Baseten released bison-1 (31B dense) and groundhog-1 (26B MoE, ~4B active), both post-trained from Gemma 4 to return calibrated probabilities rather than text. Bison-1 scores 76 on the Artificial Analysis Decision Index v2.1, ranking first of 19. Both support 128k context and images; weights are on Hugging Face, priced at $0.10 and $0.035 per 1M tokens.

    Why it matters. Open decision models with long context suit routing, triage and guardrail pipelines cheaply. Most figures are self-reported, and groundhog-1 has no official index score yet.

    ModelsImpact42Fit62Score 50

    Reported by MarkTechPost

    A niche but concrete open-weight release with pricing, context and benchmark details relevant to builders; mostly self-reported results and a narrow use case limit impact.

  22. 22

    Signal blocks screenshots on iOS and Android to keep OS AI agents from reading chats

    Signal is disabling screenshots in its iOS 27 and Android apps so assistants such as Siri and Gemini cannot read on-screen messages. Users can turn screenshots back on in settings. President Meredith Whittaker calls the fix incomplete and asks Apple, Google and Microsoft for documentation and tools to restrict AI data access.

    Why it matters. OS-level AI agents that read screens and use accessibility APIs conflict with app-level privacy. Developers currently lack proper controls, so this may push platform vendors toward permission tools for agents.

    SafetyImpact45Fit55Score 49

    Reported by The Verge

    A notable privacy conflict over OS-level agents with a concrete product change, but a partial workaround; relevant to agent builders and misuse concerns, not a model or tool release.

  23. 23

    Claude Code v2.1.330 adds enterprise sign-in, built-in chart tool, and many bug fixes

    Anthropic released Claude Code v2.1.330. It adds sign-in for organizations managed in the Enterprise Admin Console and a built-in chart tool called easel, plus new self-hosted runner options and metrics. The release also fixes dozens of bugs involving freezes, MCP, Code Mode, worktrees, and session handling.

    Why it matters. Enterprise sign-in and self-hosted runner improvements ease team deployment of Claude Code. The many MCP, subagent, and session fixes improve reliability for developers who use it daily, though this is an incremental release.

    ToolsImpact30Fit72Score 47

    Reported by Claude Code releases

    A routine point release of a major coding agent with some useful enterprise and stability changes, but nothing transformative; it is directly relevant to developer-tool interests.}

  24. 24

    Major US Banks Ban Third-Party AI Agents From Logging In, Put Liability on Users

    Wells Fargo is updating its customer terms to say use of any AI agent is at the customer's own risk. JPMorgan Chase banned third-party agents that use customer credentials, while Bank of America, Citi and PNC issued warnings or changed terms. American Express is considering reimbursement for some agent errors.

    Why it matters. Developers building consumer agents that act through logins now face access and liability barriers from banks, retailers and airlines. Unclear liability may slow adoption of credential-based agents and favor first-party or sanctioned integrations.

    PolicyImpact42Fit50Score 45

    Reported by Wired

    Several major banks restricting third-party agents is a notable deployment barrier for agent builders, but it is a terms-of-service shift, not a technical or model development.

  25. 25

    AI-driven vulnerability surge strains patching, contributing to wave of tech outages, security experts say

    Wired reports that recent outages at Microsoft, Apple, banks and the NSA were not AI attacks and had routine causes such as flawed or botched updates. Researchers say AI-assisted bug discovery and faster exploitation are pressuring teams to patch quickly. Recorded 2026 vulnerabilities passed 100,000, versus about 25,000 in 2022.

    Why it matters. AI is changing the pace of security work, shortening patch windows and raising the risk of faulty updates. Teams that ship and maintain software may need faster, better-tested deployment pipelines.

    SafetyImpact42Fit50Score 45

    Reported by Wired

    Newsletter analysis with useful figures on AI's effect on vulnerability discovery and patching; of moderate interest to AI builders and related to misuse risk, but not a model or tool release.

  26. 26

    AWS launches Amazon Textract Insights for schema-driven document extraction with confidence scores and grounding

    AWS made Amazon Textract Insights generally available. Users define fields in a JSON Schema blueprint with natural-language descriptions, then submit files up to 3,000 pages. The service splits mixed packets, matches blueprints, and returns values with confidence scores and page-level grounding. AWS also upgraded Textract Core to 40 languages and added Tokyo and Zurich regions.

    Why it matters. Per-value confidence and source grounding make LLM-based document extraction easier to audit and automate, replacing custom OCR-plus-LLM pipelines. Real-world accuracy beyond vendor claims and a customer quote is not yet shown.

    ToolsImpact38Fit55Score 45

    Reported by AWS Machine Learning

    A useful managed API launch for building document-extraction workflows, but it is an incremental vendor product with no independent evaluation; fairly relevant to developers.

Lower-ranked stories
  1. 27

    AWS Quick adds chat-built flows, multi-agent orchestration, updated web app and Meta Ads connector

    AWS released updates to Quick, its workplace AI product. Users can describe a process in chat and get a flow with deterministic steps, longer-running workflows and larger data volumes. Multi-agent orchestration runs specialized agents in parallel. The web app gains a daily briefing and background tasks, plus Meta Ads and Amplitude connectors, content safety controls and five more languages.

    ToolsImpact38Fit52Score 44

    Reported by AWS Machine Learning

  2. 28

    AWS launches Global Inference Gateway for SageMaker HyperPod, routing requests across clusters and Regions

    AWS introduced the SageMaker HyperPod Global Inference Gateway, a managed endpoint that routes inference requests across registered clusters in one or several Regions. It scores clusters on KV cache headroom, queue depth, Region proximity and error rate, and fails over automatically. Access uses SigV4, IAM and PrivateLink. AWS also announced enhanced HyperPod training observability.

    ToolsImpact35Fit55Score 43

    Reported by AWS Machine Learning

  3. 29

    UN talks on autonomous weapons end with only agreement to 'start discussing options'

    UN diplomats meeting in Geneva made little progress on rules for autonomous weapons. The Red Cross, the UN secretary-general and Pope Leo had urged binding rules within a year. The compromise gives countries two years to propose regulations, some of which may stay optional. Russia, Israel and the US were reported as main obstacles.

    PolicyImpact38Fit40Score 39

    Reported by The Verge

  4. 30

    Analysis estimates 87% of Anthropic's contracted US data center power will come from gas by 2028

    An analysis of Anthropic's compute commitments estimates about 87% of its US data center power will come from gas and 11% from coal in 2028. It puts annual CO2 emissions at 17–38 million tons and says Anthropic has made no direct US renewable commitments. Modeled health damages could reach $2–3 billion yearly.

    OtherImpact38Fit33Score 37

    Reported by Hacker News (167 points)

  5. 31

    ClawDOS: Claude-built coding agent runs natively on an IBM PC XT

    A developer released ClawDOS, an open-source coding agent running on a 4.77 MHz 8088 PC with 640K RAM. A 176 KB DOS harness talks to the Claude Messages API through a TLS bridge, has eighteen tools, compiles with Turbo C and keeps memory notes. Claude wrote all of the code.

    ToolsImpact22Fit55Score 36

    Reported by Hacker News (165 points)

  6. 32

    Meta Reconsiders 2030 Net-Zero Goal as AI Data Center Emissions Hit Record

    Meta's 2026 sustainability report says it is re-evaluating its 2030 net-zero timeline and near-term intensity targets. Total emissions rose from 15.6 million to over 21.7 million metric tonnes CO2e between 2024 and 2025, mostly from capital goods such as data center servers. Amazon, Google and Microsoft have kept their goals so far.

    BusinessImpact38Fit22Score 32

    Reported by Ars Technica

  7. 33

    Trump threatens 'tremendous tax' on companies building AI data centers outside the US

    At a Friday press conference, President Trump said he would tax companies that build data centers abroad. He gave no details and said he had not yet told his staff. The remarks come as US states and cities impose moratoriums and restrictions on new data center projects.

    PolicyImpact35Fit28Score 32

    Reported by The Verge

  8. 34

    Google employees petition leadership to drop plan to cut Search access for schools disabling Gemini

    Hundreds of Google workers, organized through the Alphabet Workers Union, are urging leadership to reverse a policy that would block schools turning off Gemini AI tools from Google Search starting next year. Parent protests took place in seven US cities. New York City, Los Angeles and Chicago school districts are discussing alternative search engines.

    BusinessImpact30Fit30Score 31

    Reported by Hacker News (282 points)

  9. 35

    Nasdaq 100 enters bear market as Nvidia falls 14% and Anthropic shares drop after IPO

    The Nasdaq 100 fell 5.6 percent on Friday and 10.4 percent for the week, its worst since the Covid crisis, leaving it 20 percent below its October peak. Nvidia dropped 14 percent on reports of an SEC accounting review, and Anthropic fell 9.1 percent a day after its market debut.

    BusinessImpact38Fit18Score 30

    Reported by The Decoder

  10. 36

    Google launches redesigned search homepage with AI buttons and no Search button

    Google is rolling out a redesigned desktop Google.com homepage after weeks of testing, confirmed by search VP Robby Stein on X. New buttons under the search bar cover image generation, file analysis and brainstorming with Gemini. The dedicated Google Search button is removed, while "I'm Feeling Lucky" stays.

    ToolsImpact30Fit28Score 29

    Reported by Ars Technica

  11. 37

    Actor licenses his face to microdrama startup Shortical for fully AI-generated series

    Shortical, a microdrama startup founded last January, premiered "Inevitable," an all-AI series using a digital avatar of actor Aki Avni, licensed under a direct deal. The company is promoting face licensing, but no platform contracts are signed and pay terms are undisclosed. Rest of World reported Chinese firms pay $15 to $700 per likeness.

    BusinessImpact25Fit22Score 24

    Reported by TechCrunch

  12. 38

    Meta's Muse AI assistant is being used to mass-unfollow Instagram accounts that don't follow back

    Wired reports that users are prompting Meta's Muse assistant to list and unfollow Instagram accounts that don't follow them back. One user had 151 accounts identified within minutes. Meta pulled an October ad promoting the unfollower-tracking feature and did not comment to Wired.

    OtherImpact18Fit22Score 20

    Reported by Wired