News · Leaderboard

Latest news, beside the leaderboard

Read what just changed, and see who is ahead. Those are the two doors on this page.

A
Ant Ling@AntLingAGI·Collected 31 min agoFresh · 6h
🚀 Release

Ant Ling Open-Sources LingSante-3.1-flash Medical Model: 560B-Total / 25B-Active MoE with 1M Context, Claims HealthBench Professional 65.72 Ahead of GPT-6 Astra, Opus 5.5 and Gemini 4 Argon

On November 22 Ant Group's Ant Ling released LingSante-3.1-flash on X, an open-weight medical model built on Ling-3.1-flash with 560B total parameters, 25B active per token and a 1M-token context window, published on Hugging Face under the MIT license. The vendor reports HealthBench Professional 65.72, HLE (text-only, with tools) 72.22 and BrowseComp (with context) 94.10, ahead of GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Argon on all three; the family also includes the 124B multimodal LingSante-3.1-air-VL.

Ant Ling Open-Sources LingSante-3.1-flash Medical Model: 560B-Total / 25B-Active MoE with 1M Context, Claims HealthBench Professional 65.72 Ahead of GPT-6 Astra, Opus 5.5 and Gemini 4 Argon
⚡ Key Takeaways
  • •Source: the official Ant Ling X account (@AntLingAGI); weights are on Hugging Face and ModelScope under the MIT license
  • •Specs: hybrid-linear MoE (63 KDA + 9 Gated MLA layers), 560B total and 25B active parameters, 512 routed experts with 8 active per token, 1M context
  • •Vendor scores: HealthBench Professional 65.72, HLE text-only with tools 72.22, BrowseComp with context 94.10, plus PhysicianBench 57 and EHR-Complex 74.40; all self-reported with no third-party replication yet
  • •Deployment: the recommended path is the SGLang day-0 image (8x H200 with TP8+EP8, or 4x GB300); vLLM is also supported and an MXFP4 quantized build is available
  • •Limits: the vendor says it is for research and evaluation only, is not a medical device and must not be the sole basis for diagnosis, prescriptions or treatment
Read details→
J
Jakub Pachocki@merettm·Collected 1h agoFresh · 6h
📊 Benchmark

OpenAI Chief Scientist Jakub Pachocki Publishes "Perplexity and Intelligence": Discloses Internal Pretraining Loss Curve and Proposes Shared Industry Limits

On November 22 OpenAI published "Perplexity and Intelligence" by Chief Scientist Jakub Pachocki, disclosing for the first time the log-perplexity of its models on an internal held-out codebase dataset since GPT-3.5: about 0.99 for GPT-3.5, about 0.91 for GPT-4, transformative cybersecurity and math capabilities below 0.8, and breakthrough math results from a model at about 0.76. He proposes perplexity as a shared cross-lab metric and common limits not to be crossed until AI monitoring reaches a concrete milestone.

OpenAI Chief Scientist Jakub Pachocki Publishes "Perplexity and Intelligence": Discloses Internal Pretraining Loss Curve and Proposes Shared Industry Limits
⚡ Key Takeaways
  • •Source: an official openai.com post by Chief Scientist Jakub Pachocki, who introduced it on X; Sam Altman reposted it calling it the most important thing OpenAI has published in a while
  • •Disclosed data: log-perplexity on an internal held-out codebase set is about 0.99 for GPT-3.5 and about 0.91 for GPT-4; recent models are below 0.8, and a model at about 0.76 began resolving open math problems
  • •Forecast: more transformative capabilities are expected in the 0.68-0.7 window, and an avalanche of pretraining algorithm advances could come as soon as 2027
  • •Proposal: frontier labs jointly set development limits based on perplexity and other measurements, not crossed until understanding and monitoring of AI behavior reaches a concrete milestone such as a successor to CoT monitors
  • •Caveats: labs use different eval sets and data mixtures, and gains from chain-of-thought reasoning do not show up in perplexity
Read details→
加
加州Will · AI時報@AIHourlyCN·Collected 2h agoFresh · 6h
📉 Price Cut

China's National Data Administration to Set Token Metering, Billing and Quality Standards; Three Token GB/T Standards Already on Its Needs List

On November 21 the National Data Administration (NDA) website said its head Liu Liehong told the 2026 World Computing Conference that China will build standards for token metering, billing and quality testing, steer "token factories" toward orderly growth and avoid low-quality duplicate buildouts. The NDA's second 2026 list of national data standards already includes three token GB/T standards (metering, multimodal token conversion, production and service quality) plus two technical documents.

⚡ Key Takeaways
  • •Source: an NDA website release on November 21 summarizing Liu Liehong's November 20 speech at the 2026 World Computing Conference in Changsha
  • •Stance: widen token service use cases, build standards for token metering, billing and quality testing, and firmly avoid low-quality duplicate buildouts
  • •Standards list: GB/T drafts for token metering, multimodal token conversion, and token production and service quality, plus technical documents on green token evaluation and token factory evaluation
  • •Progress: the WG4 group of the national data standards committee discussed each draft at its November 12-13 standards week; no release date has been announced
  • •For developers: token counting, billing and service quality for Chinese model APIs may end up under unified national standards
Read details→
T
Theo - t3.gg@theo·Collected 11h ago
🛠️ Tooling

X Discussion: Theo Says "If You Still Have Your Tech Stack in Your Bio, You're Probably Gonna Be Unemployed in 6 Months", Drawing 2,200+ Likes and 200+ Replies

On November 22, developer creator Theo (@theo, t3.gg) posted on X: "If you still have your tech stack in your bio, you're probably gonna be unemployed in 6 months." Within hours it drew over 2,200 likes, 200+ replies and about 90K views. Replies argue over whether listing frameworks still matters now that AI agents write much of the code: some say bios should list problems solved, others call the claim overblown. This is an opinion thread, not an established fact.

⚡ Key Takeaways
  • •Original post: Theo (@theo) posted at 12:17 SGT on November 22; at fetch time it had about 2,272 likes, 209 replies, 80 reposts, 36 quotes and roughly 93K views
  • •Claim: people who list a tech stack in their bio are "probably gonna be unemployed in 6 months"; the post gives no data or argument, so it is a personal view
  • •Agreeing replies: one says that with agents writing the code, the stack is the least interesting line about you; another says a bio should list problems solved, not tools used
  • •Pushback: one reply says hiring by an X bio instead of GitHub or a portfolio is a mistake, and others question Theo's view of the industry
  • •Caveat: this is an opinion thread on X, not hiring-market data
Read details→
ADSponsored
T
Tony Zhang@zhangboknight·Collected 15h ago
🛠️ Tooling

Simplex AI, Founded by Early DeepSeek Core Member Tony Zhang, Open-Sources Linger: a Local Personal Agent Whose PolyStrata Engine Runs 125B Qwen3.8-Flash-Next on a 12 GB GPU at 45.5 tok/s

On November 19, Tony Zhang, founder of Simplex AI and an early core member of DeepSeek, announced on X that Linger is now open source: a personal agent that runs on your own Mac or PC. Its MoE engine PolyStrata keeps the busiest experts on the GPU, others in RAM and rare ones on SSD; Simplex lab figures show the 125B Qwen3.8-Flash-Next writing 45.5 tok/s on a 12 GB RTX 3060 with 64 GB of RAM, with no token fees for local inference.

Simplex AI, Founded by Early DeepSeek Core Member Tony Zhang, Open-Sources Linger: a Local Personal Agent Whose PolyStrata Engine Runs 125B Qwen3.8-Flash-Next on a 12 GB GPU at 45.5 tok/s
⚡ Key Takeaways
  • •Launch: Tony Zhang (Simplex AI founder; his X bio says he was an earliest core member of DeepSeek) announced Linger as open source on X on November 19; the post has about 760K views
  • •Mechanism: PolyStrata tiers MoE weights by how often each expert is used (GPU, RAM, SSD) and promotes frequently picked experts to the GPU, so large models run on small cards
  • •Numbers (Simplex lab): the 125B Qwen3.8-Flash-Next writes 45.5 tok/s on a 12 GB RTX 3060 and 179 tok/s on an RTX 5090; the 38.1 tok/s 8 GB laptop figure was a simulated budget, not a real card
  • •Features: Model Autopilot detects hardware and installs the smartest model that fits; agents can drive local desktop apps and your signed-in Chrome and wait for approval before actions such as sending email
  • •Platforms and license: installers for macOS (Apple silicon), Windows x64 and Linux x64; the license is a modified Apache 2.0 that requires a commercial license to offer Linger as a hosted service
Read details→
A
Arena.ai@arena·Collected 20h ago
📊 Benchmark

Arena Updates Its Alignment Index: OpenAI Takes the Top 7 Spots With GPT-6.2 Sol at 85.2; Claude Fable 5.2 and Kimi K3.1 Pro Tie at 78.2, Kimi Is the Best Open Model

On November 20, Arena updated its Alignment Index with a refined method for detecting model failures. Built from 195,836 real agent sessions across 67 models, it scores unauthorized actions, deceptive completion and false attribution. OpenAI holds the top seven spots, led by GPT-6.2 Sol at 85.2; Claude Fable 5.2 is #8 at 78.2, and Kimi K3.1 Pro ties at 78.2 in #9 as the best open model, up 18.6 points on Kimi K3. Gemini 4 Hydrogen scores 71.0 and Grok 4.7 70.8.

Arena Updates Its Alignment Index: OpenAI Takes the Top 7 Spots With GPT-6.2 Sol at 85.2; Claude Fable 5.2 and Kimi K3.1 Pro Tie at 78.2, Kimi Is the Best Open Model
⚡ Key Takeaways
  • •Method: the Alignment Index measures three agent failures in real user sessions: unauthorized action (doing what the user did not ask for or approve), deceptive completion (claiming work is done or checked when it is not) and false attribution (crediting the user with something they never said)
  • •Ranking: OpenAI holds the top seven, with GPT-6.2 Sol at 85.2, GPT-6 Astra at 84.6 and GPT-6.1 Terra at 84.5; the leaderboard page is dated November 19 and covers 195,836 sessions and 67 models
  • •Anthropic: Claude Fable 5.2 debuts at #8 with 78.2; Arena says Claude Opus 5.5 fell from #6 on the previous index to #17
  • •Open models: Kimi K3.1 Pro is #9 at 78.2, up 18.6 points and 28 places on Kimi K3; the next best open model, MiMo-V2.6-Flash, is only #26
  • •Also: Gemini 4 Hydrogen scores 71.0, Grok 4.7 70.8 and Gemini 4 Neon 70.4; OpenAI president Greg Brockman shared the results on X as "OpenAI models topping the alignment benchmarks"
Read details→
T
Tesla@Tesla·Collected 23h ago
🚀 Release

Tesla Rolls Out FSD v15: Every Part of the End-to-End Network Upgraded, 25x Better on Safety-Critical Cases; Musk Says the New In-Car View Is Real-Time AI Video on Half an AI4 Chip

On November 21, Tesla announced on X that FSD v15 is starting to roll out, upgrading every part of its end-to-end driving network, with 25x better handling of safety-critical cases and a new tap-to-go Tesla Reality View. Elon Musk later said the new in-car view is real-time AI video generation, with no game engine, running on half of one AI4 accelerator, and Tesla AI lead Ashok Elluswamy said it uses less compute than the old visualization.

⚡ Key Takeaways
  • •Release: Tesla's official X account says FSD v15 is starting to roll out, upgrading every part of the end-to-end driving network; rollout pace and eligible vehicles were not stated
  • •Safety: Tesla claims 25x better handling of safety-critical cases, a company-stated figure with no test method or third-party data attached
  • •Reality View: a live view that shows what FSD plans to do and lets drivers control it; Elon Musk says you can tap any point to make the car go there
  • •Real-time generation: Musk says the new view is real-time AI video generation with no game engine, using half of one AI4 accelerator; Ashok Elluswamy says it uses less compute than the old visualization; neither shared model details
  • •Also: trip start and end now learn end-to-end from fleet route preferences; better detection of signs, animals, potholes and standing water, slowing for speed cameras, and a new Bike Rack Mode
Read details→
E
Elon Musk@elonmusk·Collected 1d ago
🛠️ Tooling

X Discussion: Elon Musk Calls SFT "a Haircut with a Chainsaw" and System Prompts "Silly", Post Draws 14K Likes

On November 21, Elon Musk posted on X that SFT (supervised fine-tuning) is "more like getting a haircut with a chainsaw" than fine-tuning and should be built into primary training, and that having a system prompt is "silly" because default behavior should be sensible. He said both are already true for Tesla AI. At writing time the post had about 14,600 likes, 1,241 replies and about 2.2 million views. This is an opinion, not a technical paper.

⚡ Key Takeaways
  • •Point one: Musk argues SFT is a crude after-the-fact edit, "a haircut with a chainsaw", and should be built into primary training
  • •Point two: he calls having a system prompt "silly", saying default behavior should be sensible
  • •He says both are already true for Tesla AI, calls it "the most efficient AI on Earth" and says FSD V15 goes to wide release next month; these are his claims, with no data attached
  • •Engagement at writing time: about 14,600 likes, 1,188 reposts, 1,241 replies, 391 quotes and about 2.2 million views
  • •Pushback: one quote post notes that for an auditor the system prompt may be the only readable statement of agent policy, and moving it all into training leaves governance with no document to inspect
Read details→
ADSponsored
J
Julien Chaumond@julien_c·Collected 1d ago
🛠️ Tooling

Hugging Face Runs Its Own CDN for AI Downloads: Nearly 1 Exabyte of Models and Datasets per Month, About 420 GB/s and 80,000 Requests per Second

On November 20, Hugging Face co-founder and CTO Julien Chaumond said on X that Hugging Face serves almost 1 exabyte of open models and datasets over the internet per month and had to build its own CDN for AI downloads, sharing a live traffic map. Co-founder Thomas Wolf added that the CDN has run since this summer and handles about 420 GB per second through 80,000 requests per second.

⚡ Key Takeaways
  • •Scale: Hugging Face serves almost 1 exabyte (about 1 billion GB) of open models and datasets per month, as disclosed on X by co-founder and CTO Julien Chaumond
  • •Throughput: co-founder Thomas Wolf says this is about 420 GB per second through 80,000 requests per second, and the in-house CDN has run since this summer
  • •Why build one: classic CDNs were designed for many small web files sent to human users across many cities; AI downloads are a few huge files, largely pulled by machines, often GPU clusters in a handful of cloud regions
  • •Design focus: caches tuned for massive objects that can absorb release-day floods and several giant datasets and trending models being pulled at once
  • •The team published a live traffic map of where models and datasets flow; all figures are Hugging Face-stated, with no third-party audit attached
Read details→
H
House Foreign Affairs Committee Dems@HouseForeign·Collected 1d ago
🚀 Release

House Foreign Affairs Committee Discloses BIS China AI Chip Licenses: 879,868 H200 and MI325X Approved, All Marked Approved, Not Shipped as of May

On November 19, House Foreign Affairs Committee Chairman Brian Mast used his authority under the Export Control Reform Act to publish BIS license information for Nvidia H200 and equivalent AMD chips bound for China: about 3.4 million chips were requested, and as of May 11, 2026, 21 applications covering 879,868 chips were approved, every one marked "Approved, not shipped". Approved recipients include Alibaba, ByteDance, Tencent, Baidu and JD.

⚡ Key Takeaways
  • •Totals: Nvidia and AMD requested to sell about 3.4 million AI chips to China; as of May 11, 2026, BIS approved 21 applications covering 879,868 chips
  • •Split: adding the published rows gives 579,868 approved Nvidia chips (17 entries) and 300,000 AMD MI325X chips (4 entries of 75,000 each)
  • •Recipients: Alibaba 75,000, Tencent and subsidiaries 150,000, ByteDance and subsidiaries 150,000, Baidu 43,000, JD 41,300, miHoYo 57,810 and others
  • •Status: every approved entry is marked "Approved, not shipped"; other applications were returned without action (RWA), still pending, or rejected (UCloud, 75,000 chips)
  • •Context: the committee called it the most detailed public disclosure in the history of the Export Control Reform Act and pledged oversight of the administration's export control policy
Read details→
M
Michael Luo@AzianMike·Collected 1d ago
🛠️ Tooling

OpenAI Agents API Now Supports Daybreak: Session-Level access_programs.cyber Selects Blue or Red for Always-On Security Agents

On November 20, Michael Luo, who works on product at OpenAI, announced on X that the Agents API now supports OpenAI's Daybreak cybersecurity access program. Official docs show developers set access_programs.cyber (daybreak_blue or daybreak_red) when creating or updating a session rather than on individual messages; Blue pairs with general models such as GPT-6 Sol, Red unlocks GPT-5.6 Cyber, and GPT-6 Cyber needs additional approval. Organizations must already be approved for Daybreak and enable it in a named, non-default project.

⚡ Key Takeaways
  • •Usage: add access_programs.cyber to the request body when creating or updating an Agents API session, set to daybreak_blue or daybreak_red; it cannot be set per message
  • •Models: the Blue example uses GPT-6 Sol and GPT-6 Luna also qualifies; the Red example uses GPT-5.6 Cyber; GPT-6 Cyber needs Red plus additional model approval
  • •Prerequisites: the organization must be approved for Daybreak, an org owner enables it in a named (non-default) project, and requests use that project's API key
  • •Compliance: every Daybreak Red organization must require SSO for the API organization holding Cyber access and enforce MFA through its identity provider
  • •Use cases (listed by Michael Luo): investigating suspicious activity, implementing and verifying security fixes, triaging for human review, and quarantining suspicious actors
Read details→
H
HD Moore@hdmoore·Collected 1d ago
🛠️ Tooling

Dragos Seeds About 2 Million Canary Prompts and Finds Three Compromised Providers and Prompt-Reading Resellers; Novita and Together AI Follow With Incident Notices, Together Blaming a Credential Committed by Its Coding Agent

On November 19, HD Moore, Chief Research Officer at security firm Dragos, published research in which his team sent about 2 million prompts containing unique URLs and decoy AWS keys to widely used AI services and routers between July and November 2026, then tracked where those canaries surfaced. They found at least three inference providers compromised by attackers (the bait AWS keys were used for tens of thousands of API calls), marketplace resellers viewing and keeping customer prompts through built-in logging, and providers ignoring privacy settings: one Chinese AI lab fetched secret URLs from more than half of the prompts sent to its models from July to October, including roughly two-thirds of requests marked Zero Data Retention. On November 21, Novita published a privacy notice confirming some MiniMax requests had been forwarded to MiniMax's official API. The same day, Together AI disclosed that a credential committed by an AI coding agent let an attacker download prompt samples from June and July, including Dragos canary credentials.

⚡ Key Takeaways
  • •Scale: about 2 million prompts from July to November, each carrying a unique URL, DNS name and decoy AWS key, traceable to the exact model, route and privacy setting
  • •Breaches: at least three inference providers were compromised, the earliest from mid-July; attackers copied requests at the front door, and bait keys drove tens of thousands of AWS API calls, later including the ARTEX AI pentest framework
  • •Snooping: individual marketplace resellers only had to list a model at the lowest price and turn on request logging to read customer prompts; the leak followed the seller, not the whole marketplace
  • •ZDR failures: one Chinese lab fetched secret URLs from more than half of prompts (about two-thirds of ZDR requests); one link appeared in a Feishu chat 24 days later
  • •Follow-up: Novita confirmed some MiniMax requests from May to November, including those routed via OpenRouter, were forwarded to MiniMax's official API; Together AI disclosed that a credential committed by its coding agent let an attacker download June-July prompt samples for MiniMax M3, GLM-5.2 and Kimi K2.6, with five of the tested credentials being Dragos canaries, and now defaults new accounts to ZDR
  • •Advice: self-host or go direct to a first-tier provider, pin upstream providers on routers, avoid reseller marketplaces, seed your own canaries in agent context and enable secret filtering
Read details→
S
Signal@signalapp·Collected 1d ago
🛠️ Tooling

Signal Restricts Screenshots by Default to Block OS-Level AI Agents: Blocked on iOS 27, Auto Screen Security on Android When an AI Assistant Reads the Screen

Signal announced on its official blog on November 20 that, to stop agentic AI assistants built into mobile operating systems (such as Siri AI onscreen awareness on iOS 27 and Gemini Intelligence on Android) from capturing and storing chats, it now restricts screenshots in the app by default: screenshots are blocked on iOS 27, and on Android the Screen security setting turns on automatically when an AI assistant tries to read the screen. Users can re-enable screenshots in settings. President Meredith Whittaker called OS-level AI agents an existential threat to privacy and urged Apple, Google and Microsoft to give developers a simple, documented way to opt apps out of AI data access.

Signal Restricts Screenshots by Default to Block OS-Level AI Agents: Blocked on iOS 27, Auto Screen Security on Android When an AI Assistant Reads the Screen
⚡ Key Takeaways
  • •New default: Signal blocks in-app screenshots on iOS 27; on Android it auto-enables Screen security when it detects an AI assistant reading the screen
  • •Targets: OS-level agents such as Siri AI onscreen awareness and Gemini Intelligence, plus third-party agents like Meta's Muse that use accessibility permissions
  • •Limits: Signal admits the fix is incomplete and cannot stop filesystem access; macOS lacks a reliable API, so users should add Signal to Siri's Excluded Apps
  • •Ask: OS vendors should offer a documented opt-out from AI data access covering screen reading, filesystem, notifications and APIs, plus clear specs on how agents process data
  • •Opt-out: users who want screenshots can disable the protection, at the cost of exposing their own and their contacts' messages to AI agents
Read details→
不
不明真相的群众@__jingling·Collected 1d ago
📊 Benchmark

MiniMax Reveals Automated Cryptanalysis by Its Internal MIM Model: First Practical Breaks of Several NIST Candidates, Waterfall-256 Collisions in 0.045s and Up to 43,400x Faster McEliece Solving

MiniMax said on its official blog on November 20 that its research team used an internal-only research model, MiniMax's Internal Model (MIM), for automated cryptanalysis. On CryptanalysisBench Tier 2 it produced the first practical breaks of several NIST competition candidates that had no known practical attacks, found new attacks on peer-reviewed or award-winning designs such as COLM, SBC-VOLE and the NSA's ARADI, and solved 55 previously unsolved TII McEliece Challenge instances. MiniMax stresses the work covers academic designs, past candidates and public challenge parameters and does not affect deployed standards. Technical reports and proof-of-concept code are on GitHub; the model itself is not public, and researchers can apply for test access.

MiniMax Reveals Automated Cryptanalysis by Its Internal MIM Model: First Practical Breaks of Several NIST Candidates, Waterfall-256 Collisions in 0.045s and Up to 43,400x Faster McEliece Solving
⚡ Key Takeaways
  • •NIST candidates: collisions on SHA-3 candidate Waterfall-256 across all output lengths in 0.045s on one core; full-round Blender-256 collisions in 0.36-2.3s
  • •Post-quantum: LOCKER solver matrix shrunk from 42 PB to 58 MB with the shared secret recovered in 2s; ROLLO-I-128 equivalent secret key from the public key in about 6s
  • •Implementation bugs: more than a dozen unreported flaws in reference code, e.g. DRS mistyped && as &, enabling signature forgery with zero signing queries
  • •Peer-reviewed designs: COLM forgery from two encryption queries; SBC-VOLE full key from one signature in 0.06s; ARADI practical key recovery pushed from 5 to 8 rounds
  • •Caveats: results are self-reported and the model stays internal; MiniMax says deployed standards are unaffected and scheme authors were notified in advance
Read details→
S
StatsWire@StatsWire·Collected 1d ago
🛠️ Tooling

Gemini API Now Accepts Word, PowerPoint, Excel and OpenDocument Files Directly and Lets Developers Slice Pages, Sheets, Cell Ranges and Audio/Video Intervals, Billing Only the Slice

Google's official Gemini API changelog for November 18 announced multimodal file processing and slicing: Microsoft Office and OpenDocument files (Word, PowerPoint and Excel) can now be passed to the Gemini API directly, and documents, spreadsheets, videos and audio can be sliced in user prompts and function responses. Developers pick page ranges, worksheet tabs, cell ranges, time intervals or content layers such as page images, extracted text, formulas or video transcripts, and only process and pay for that content.

⚡ Key Takeaways
  • •Since November 18 the Gemini API accepts Office and OpenDocument files such as .docx, .pptx, .xlsx, .doc/.ppt/.xls and .odt/.odp/.ods directly, processed the same way as PDFs
  • •Add processing (type static) on a document block to slice: page_options sets start/end pages and content layers (page_image / page_text), spreadsheet_options sets sheet_index and an A1 cell range such as A1:H50
  • •With processing set, only the selected pages, sheets and content layers are counted and billed; page images cost about 280 / 560 / 1,120 tokens per page by resolution, medium by default
  • •Video can be clipped with start_offset / end_offset and limited to frames, audio or transcript layers, audio supports interval clipping, and tools can return slices inside function results
  • •Limits: 50 MB per file, up to 2,000 pages or slides per Office file and 50,000 cells per sheet, and the limits still apply to the whole file when slicing
Read details→
G
Google@Google·Collected 1d ago
🛠️ Tooling

Gemini API Gets Voice Editing: Gemini 3.5 Transcribe Smart Mode Edits, Rewrites and Translates Existing Text From Spoken Instructions, the Tech Behind Rambler on Gboard, in Batch and Live Transcription

Google said on X that it is bringing Gemini 3.5 Transcribe's voice editing, the same technology behind Rambler on Gboard and the Gemini app on macOS, to the Gemini API. The official changelog shows gemini-3.5-transcribe and gemini-3.5-transcribe-live gained voice-driven editing, rewriting and formatting on November 17. It is on by default in Smart transcription mode: users can dictate and give spoken edit instructions in the same recording, or pass up to 64 KiB of existing text in source_text, and the model applies the instructions and returns the complete updated text.

⚡ Key Takeaways
  • •gemini-3.5-transcribe and gemini-3.5-transcribe-live in the Gemini API gained voice editing on November 17; Google promoted it on X on November 20
  • •It is on by default when transcription_config mode is smart, so dictation and spoken edit instructions can be in the same recording
  • •Pass up to 64 KiB of existing text in source_text and the model edits, rewrites or formats it from spoken instructions, returning the complete updated text in output_text
  • •In Live API sessions the server keeps the active text and emits the full replacement in input_transcription; interim_input_transcription only streams the literal instruction, so wait for the final result before updating a document
  • •Paid-tier pricing is $2.00 per 1M audio input tokens and $12.00 per 1M text output tokens, an estimated blended ~$0.005 per minute; no voice-editing benchmarks were published
Read details→
A
Andrew Ambrosino@ajambrosino·Collected 1d ago
🚀 Release

ChatGPT Space Rolls Out to Plus: Shared Interactive Pages and Slides That Auto-Update From Connected Apps, Replacing Library; OpenAI's Andrew Ambrosino Says a Redesign and Rebuilt Threads Ship Alongside

OpenAI's ChatGPT release notes for November 20 say ChatGPT Space is rolling out globally to Plus users aged 18 and older, after launching for Pro, Business and Enterprise. In Space you create interactive pages and presentations with ChatGPT, invite others with Space access to co-edit them, and set pages and slides to update automatically from connected apps and tools; Space replaces Library in the sidebar. Andrew Ambrosino, who works on Codex and ChatGPT at OpenAI, said on X that the same release brings a simpler visual design, rebuilt threads, much faster pages and slides, page editing on mobile and more reliable Codex.

⚡ Key Takeaways
  • •Starting November 20, Space is rolling out globally to Plus users aged 18 and older; it was previously limited to Pro, Business and Enterprise
  • •ChatGPT can create interactive pages and presentations, and pages and slides can update automatically from connected apps and tools, such as a self-updating to-do list or a project plan that tracks decisions from Slack
  • •Tag ChatGPT in a comment to request changes; opening a shared item requires a Plus, Pro, Business or Enterprise account with Space access, Free users cannot open them, and there are no public links
  • •Space replaces Library in the sidebar as the home for pages, slides, images and uploaded files, on web, desktop, iOS and Android
  • •Andrew Ambrosino said on X that the same release adds a simpler visual design, rebuilt threads, faster pages and slides, page editing on mobile and more reliable Codex (no performance figures published)
Read details→
G
Google Antigravity@antigravity·Collected 1d ago
🛠️ Tooling

Google Antigravity Upgrades /btw to Fork, Explore and Promote Side Threads; App v2.34.0 Brings Gemini 4 Neon to Paid Consumer Plans and MCP Apps, Varun Mohan Says Bigger Neon Cache Pool Stretches Quota at Least 20% With a Reset Coming

Google Antigravity announced on November 20 that /btw now does much more than one-off side questions: instead of a quick question while the main agent is busy, it lets you fork, explore and promote side threads without losing your place, including in the CLI. App v2.34.0, released the same day, brings Gemini 4 Neon and Orchestrator (Neon) to eligible paid consumer plans and adds MCP App support. Separately, Antigravity lead Varun Mohan said Google is massively enlarging the Neon cache pool, raising cache hit rate about 4% and stretching quota at least 20%, with a quota reset once the rollout is done.

⚡ Key Takeaways
  • •/btw forks the current conversation: the copy keeps full context, the main agent keeps running, and nothing in the fork is written back to the main conversation
  • •In the App, a side question opens as a Side Question tab that supports multi-turn follow-ups and can be promoted to a conversation or deleted; in the CLI each /btw is a separate fork card and tool calls that need approval are denied automatically
  • •App v2.34.0 (November 20) brings Gemini 4 Neon and Orchestrator (Neon) to eligible paid consumer plans and to Gemini API keys, adds MCP App support, and lets you start a side question from selected terminal output
  • •Varun Mohan said on X that a much larger Neon cache pool raises cache hit rate about 4% and stretches quota at least 20%; because the old behavior burned quota faster, a quota reset will follow once the rollout is done (no date announced)
Read details→
M
METR@METR_Evals·Collected 1d ago
📊 Benchmark

METR Q4 2026 Capability Update: Frontier Time Horizons Double About Every 3.6 Months, Naive Extrapolation Reaches 200-Hour Tasks by July 2027, AI R&D Capabilities Getting Harder to Measure

AI evaluation nonprofit METR published a research update on November 20 covering public models up to Opus 5.5 and GPT-6 Astra. On its expanded TH1.1 plus TH2 task suite, frontier 50% time horizons double roughly every 3.6 months (95% CI 2.1 to 4.9), and naive extrapolation reaches the suite's longest 200-hour tasks in July 2027. METR says current public evidence is insufficient to rule out drastic AI R&D acceleration in the next couple of years, and that capabilities are getting harder to measure.

⚡ Key Takeaways
  • •On the combined TH1.1 and TH2 suite, frontier 50% time horizons double roughly every 3.6 months (95% CI 2.1 to 4.9), close to the trend OpenAI reported on internal research tasks
  • •Naive extrapolation has 50% time horizons reaching the suite's longest tasks (200 hours) in July 2027
  • •On coding and optimization tasks, models usually match or beat the skilled humans METR recruited; on contest-style tasks with many entrants they place very high but not at the top
  • •The new tasks are so hard that only 9% of human time estimates come from successful human baselines; the rest use records, expert interviews and extrapolation
  • •METR introduces the METR Capability Index (MCI), which, like Epoch's ECI, estimates ability directly from scores; trend fits range from slight deceleration to moderate acceleration depending on methodology
Read details→
H
House Homeland GOP@HomelandGOP·Collected 1d ago
🛠️ Tooling

House Committees Release Joint Report on PRC Open-Weight Models: Cursor Maker Anysphere Plans to Explore Phasing Them Out, Composer 2 Built on a Moonshot Base; Airbnb Runs Qwen, DoorDash Deployed 25 PRC Model Families

On November 20, House Homeland Security Committee Chairman Andrew Garbarino and House Select Committee on China Chairman John Moolenaar released a joint investigative report on U.S. companies using PRC-developed open-weight models, with Airbnb, Anysphere (Cursor) and DoorDash as case studies. It says Cursor's flagship Composer 2 model was built on a Moonshot AI base model and that Anysphere told the Committees it plans to explore phasing out PRC-developed open-weight models because of geopolitical risk. The report makes 12 recommendations, including prohibiting PRC models in federal systems and high-risk critical infrastructure.

⚡ Key Takeaways
  • •The report says Anysphere built its flagship Composer 2 model on a Moonshot AI base model, which became public only after an independent developer identified it, and that Anysphere told the Committees it plans to explore phasing out PRC-developed open-weight models because of geopolitical risk (no timeline announced)
  • •Airbnb runs customized Alibaba Qwen models on U.S.-based provider infrastructure for about 80,000 support interactions and about 250,000 case summaries per day
  • •DoorDash reported evaluating models from at least six PRC developers, with 25 PRC model families deployed in a non-trivial way; replacing one embedded model could take up to six months
  • •Per the report, PRC models made up 41 percent of downloads on the leading open-model repository over the past year, and the top American open-weight model ranked tenth, behind six PRC models
  • •The 12 recommendations include a national American open-weight AI initiative, a model provenance standard and AI bill of materials, banning PRC models in federal systems and contract work, and holding illicit distillers accountable
Read details→