Stay ahead in AI without the scroll.

10 minutes a day: what changed, why it matters, and what to try next. Free, every morning, 7 days a week.

A steaming white teacup and saucer beside a brass service bell on a weathered wooden windowsill, with a misty town street at sunrise outside

· 19 min read

OpenAI lets businesses claim ChatGPT profiles; Gemini 4 Carbon opens to developers

Published

OpenAI launched Businesses in ChatGPT, letting firms claim a profile, see mentions and run their own agent. Google opened Gemini 4 Carbon to every paying developer, and a security firm documented an AI agent that spent two weeks inside a European telecom.

Today in 30 seconds

  • OpenAI launched Businesses in ChatGPT in beta: claim a listing, see mention and click counts, and (for eligible US firms) run a Business Agent inside ChatGPT.
  • Google opened Gemini 4 Carbon to every paying developer, with a /swarm preview and outputs up to 1M tokens.
  • Malbear Labs found transcripts of a jailbroken DeepSeek Flash agent that spent two weeks inside a European telecom.
  • Anthropic briefly emailed banks a filing for an IPO of up to $69.3B, then said the pricing details were wrong.
  • On a new visual knowledge-work bench, Fable 5.2 leads at 57%, with Gemini 4 Carbon close behind at a fifth of the cost.

Pour a slow cup today: the biggest change isn’t a model, it’s a shop window. OpenAI now lets any business claim its profile inside ChatGPT, see how often it gets mentioned, and put its own agent in front of 1.2 billion weekly users.

Google opened Gemini 4 Carbon to every paying developer, Anthropic’s IPO paperwork had a very public wobble, and a security firm found the full workspace of an AI agent that spent two weeks breaking into a European telecom.

Top news

Platforms

OpenAI lets businesses claim their ChatGPT profile, see their mentions and run their own agent

OpenAI launched “Businesses in ChatGPT” in beta on 10 November. Through a new Business Hub, any eligible business can claim and verify its listing, edit its details, upload product feeds and see how often ChatGPT mentions it. Eligible US e-commerce and online-service businesses can also build a Business Agent that talks to customers inside ChatGPT.

  • Claiming: Any eligible business in a country where ChatGPT is available can claim itself in Business Hub.
    • Once verified, it can update key details and add self-service product feeds with pricing and availability.
    • Shopify merchants can connect from Business Hub or the Shopify admin, and their eligible products are already in ChatGPT through Shopify Catalog.
  • Insights: Business Hub shows aggregated counts of mentions and website clicks, plus “discovery topics”, meaning the common themes in conversations where the business’s site gets cited.
    • OpenAI says it shares no user conversations or Business Agent chats, and insights only appear above minimum volume thresholds.
  • Business Agents: When a business appears in an answer, users can open its ChatGPT profile and, if it has one, start a separate, clearly labelled chat with its agent.
    • Businesses can sponsor that agent and use it as the landing spot for ChatGPT Ads; sponsored placements are labelled and kept apart from ChatGPT’s own answers.
  • Scale: OpenAI puts ChatGPT at 1.2 billion weekly users. OpenAI’s Adam Fry wrote that it went from 1B to 1.2B “in just a few months”.
  • Reaction on X: @gregisenberg called it a possible “Google My Business moment for AI”, recalling how agencies made millions in the 2010s claiming and fixing listings.
    • SEO accounts were already posting step-by-step claim guides within hours.

Why it matters: This is the first official view of how a brand shows up in AI answers, the thing the “AI search optimisation” industry has been trying to guess from the outside. It also opens a paid channel inside ChatGPT, so expect the line between ChatGPT’s own answer and a business’s sponsored agent to become a trust question quickly.

  • Do this now: If you run a site or shop that people might ask ChatGPT about, claim it in Business Hub and check what it already says about you, before someone else defines it.

Sources: OpenAI announcement · Reuters · @OpenAI · @gregisenberg

Models

Gemini 4 Carbon opens to every paying developer, with /swarm teams and 1M-token outputs

Two weeks after it went only to cyber defenders, Google made Gemini 4 Carbon available on 10 November in the Gemini API (paid tier), AI Studio, Gemini for business, and for Google AI Ultra subscribers in Antigravity, the Gemini app and AI Mode in Search.

It comes with a new /swarm preview in Antigravity that turns one prompt into a coordinated team of agents.

  • Access and price: Model id gemini-4-carbon, default thinking_level medium.
    • The introductory price matches Argon at $2 / $10 per million input / output tokens until 30 November, then $4 / $20 from 1 December, with 95% off cached input.
    • Switching from Argon is a one-string change.
  • Long outputs: Input, thinking and output share one 1M-token window. Output defaults to 64K tokens; with background=True and a higher max_output_tokens it can write far past that, up to 1M minus the input.
    • Google tells developers to raise client timeouts to 15–20 minutes for hard prompts.
  • Migration notes: thinking_budget is deprecated in favour of thinking_level; temperature, top_p and top_k are ignored; explicit context caching gives way to automatic implicit caching.
  • Benchmarks Google cites: #1 on CursorBench 4.0, Perplexity’s WANDR and Zapier’s AutomationBench at “more than 2X lower cost per task” than the next external model.
    • Cursor says it scores 60.4% on CursorBench at high effort.
  • Safety numbers: In simulated attack environments Carbon blocked 99% of adaptive red-team attack paths, with a per-attempt breach rate of 0.8% (Argon: 3.1%). On a 375-task agentic honesty test it passed 92% (Argon: 70%).
  • Science claims: Google says Carbon helped researchers formally prove security properties of Signal’s protocol code in Lean (with a new AlphaProof Code framework), and retrieved 40× more remote protein homologs on average for 7 low-homolog CASP17 families.
  • Early reviews: @emollick called it “very good” and useful because it “thinks differently than Claude/GPT”.
    • @theo said he was underwhelmed at first, then wrote that after more tests “this video will not be going how I initially planned.”
    • @_overment reported Carbon failing on every query in Cursor on launch day.

Why it matters: The cyber-first gate is over, so the strongest Gemini model is now a normal API call at Argon’s price, at least until December. The pairing Google recommends, Carbon as planner and checker with cheaper Flash or Argon workers, is the multi-agent pattern most teams are moving towards anyway.

  • Do this now: If you already call Argon, run your hardest eval set on gemini-4-carbon before 30 November, while it costs the same.

Sources: Google blog · Developer guide · @OfficialLoganK · @cursor_ai via @demishassabis · @theo

Security

Operation Snowball: a jailbroken DeepSeek Flash agent spent two weeks inside a European telecom

Malbear Labs published research on 8 November after finding an attacker’s open, unprotected working directory. It held more than 70 session transcripts showing an AI agent running largely unsupervised for over two weeks against a European telecom provider and 13 organisations hosted in its data centres and cloud.

  • The setup: A custom agent framework (version 0.2.0-rc.2) built on DeepSeek Flash with a 1M-token context, so the lead agent could hold the whole state of a multi-day intrusion and hand tasks to worker agents.
  • The jailbreak: A system prompt of about 78,000 characters, built on a public jailbreak template, told the model to treat any refusal as a “malfunction”. Malbear says the transcripts contain no recorded refusals.
  • The human: The operator wrote in Chinese, often just “Go on” or “Stop”, set target subnets and marked three hosts off-limits.
    • Malbear says that is consistent with a Chinese-speaking operator rather than a state actor; language alone is not attribution.
  • How it got in: CVE-2026-20320, an unauthenticated XXE flaw in Cisco BroadWorks that Cisco fixed in August, then old Liferay (CVE-2020-7961) and JBoss/RichFaces (CVE-2013-2165) bugs for deeper footholds.
  • It adapted on its own: When endpoint security blocked a PwnKit exploit, the agent changed approach without being told; when a tunnelling tool was blocked, it wrote a replacement.
  • Claimed haul: Dozens of credential sets, private keys, a healthcare organisation’s Active Directory data, subscriber billing records and, per the agent’s own reports, millions of patient records through an Oracle database still on default credentials.
    • Malbear’s post does not say these claims were confirmed with any victim, and no victims are named.

Why it matters: This is one of the clearest inside views yet of an autonomous AI intrusion, and nothing in it was new: every bug used was already patched, some for over a decade. The agent’s advantage was patience and scale, which means the boring defences (patching, no default passwords, watching for unusual tunnelling) matter more, not less.

  • Do this now: Check for internet-facing end-of-life servers and any database still on default credentials; those are exactly what this agent harvested.

Sources: Infosecurity Magazine · PlayCISO teardown · CISO Voice

Markets

Anthropic briefly sent banks a $69.3B IPO plan, then said the pricing was wrong

Late on 9 November an Anthropic representative emailed banks a securities filing showing about 867 million shares at $70–$80, which would raise up to $69.3 billion. A follow-up email said the pricing details were inaccurate and an updated filing would come later, Reuters reported from emails it saw.

  • What the retracted filing showed: 866.7 million Class A shares, about 843.3 million sold by Anthropic and 23.3 million by existing holders.
  • A $26 billion side deal: Google, Nvidia, AMD, Broadcom, Micron, Microsoft and TSMC had agreed to buy Class N shares in a private placement at the IPO price, closing right after the listing.
  • What’s not known: The second email didn’t say what was wrong, and Reuters couldn’t determine it. Anthropic declined to comment.
    • Bloomberg Law reported the company later called the number outdated, and noted it fell short of earlier hopes of matching SpaceX’s record debut.
  • Other figures in reporting: The Times of India, citing the disclosures, reported a Nasdaq listing under the ticker ANT, first-half 2026 revenue of $16.3 billion and a net loss of $80.7 billion.
    • Much of the loss is from fair-value changes on financing instruments rather than cash spending.
  • Mood: It lands while other AI infrastructure and data-centre listings have been delayed or pulled amid higher interest rates and a selloff in names like CoreWeave.

Why it matters: The roadshow is under way and pricing is still pencilled in for the coming week, so the corrected filing is the number to watch. The $26 billion from chip and cloud suppliers is a reminder of how circular AI financing has become: the companies selling Anthropic compute are also buying its stock.

Sources: Reuters via Investing.com · Bloomberg Law · Times of India

Benchmarks

New MKA test: Fable 5.2 leads visual knowledge work, but Carbon gets close for a fifth of the cost

Google DeepMind and Artificial Analysis launched MKA (Multimodal Knowledge Agent), 200 private tasks where an agent must read dense visual files such as architectural drawings, schematics and clinical scans, then produce real deliverables like decks, spreadsheets and marked-up images.

  • Format: 200 tasks across 10 industries and 83 occupations, written by domain experts.
    • Each is graded on about 70 weighted rubric criteria, split into baseline “Regression” checks and harder “Climb” checks; the headline score is the share of Climb criteria met.
  • Harness: Runs in Artificial Analysis’s open-source Stirrup harness with code, web search, fetch and image-viewing tools for up to 250 turns, graded by GPT-6 Luna as an agentic judge.
  • Results: Claude Fable 5.2 (max) 57%, Gemini 4 Carbon (high) 54%, Claude Opus 5.5 (max) 53%.
  • Cost: Fable averages $31.97 per task, over 5× Carbon’s $6.29 for a few extra points, and about 29× GPT-6.1 Sol ($1.11).
    • All four OpenAI models tested sit on the cost-performance frontier below Carbon.
  • No single winner: Fable leads five verticals, Opus 5.5 leads Education and Construction, Argon leads Science, Carbon leads Healthcare and Sonnet 5.5 narrowly leads Law & Government.
    • Word and plain-text outputs are the hardest; no model passes 52% on either.

Why it matters: It’s one of the few evals that requires reading the picture rather than the text around it, which is what most office work with drawings and scans actually needs. The cost spread is the useful bit: within a few points of the top, prices range nearly 30×.

Sources: Artificial Analysis article · MKA leaderboard · @_philschmid

Products

Claude can now open your full apps beside the chat, and every paid user gets a free limit reset

Anthropic’s 10 November update lets supported connectors open the full version of a tool, such as Adobe, Atlassian or Box, beside a Claude conversation or full screen, where you and Claude edit it together. Claude Data (formerly Dashboards) is now generally available, and Claude Sheets and Whiteboards arrived in beta.

  • Interactive connectors: Rolling out on web and desktop for accounts with the new Claude experience, not yet on mobile.
    • Examples from Anthropic: retouching photos and redacting PDFs in Adobe, editing a Confluence page live, revising a Box file and comparing versions before keeping changes.
  • For developers: These run on MCP Apps, the open MCP extension that lets a server bring its own interface. Claude now supports more of it, including app tools and embedded sites, so a whole app can run inside Claude.
    • That support is a developer public preview, and an MCP App isn’t tied to Claude.
  • New work tools: Claude Sheets (shared spreadsheets with formulas and tabs) and Whiteboards are in beta.
    • Claude Data connects to tools like Snowflake and Salesforce and drafts dashboards, data stories and narrated videos for review.
    • TechRadar reports Claude Meetings is coming too. Salesforce (beta), NetSuite and Workday (beta) plugins ship with built-in skills.
  • The reset: @ClaudeDevs says all Pro, Max, Ultra and Team users got one free full reset across all models and limits, usable any time before 10 December at 2:00 PM PT (23:00 Warsaw time).
    • Anthropic says it fixed “a few small issues that used more limits than intended”.

Why it matters: Claude is moving from “assistant that calls tools” to a workspace where the tools themselves live, the same direction OpenAI is taking with apps. For builders, MCP Apps is now the shared way to ship a UI into more than one AI client.

  • Do this now: If you hit your Claude limit mid-week, the free reset is a one-off; save it for a heavy Fable session rather than spending it by accident.

Sources: Anthropic: open your apps in Claude · Anthropic: Sheets, Whiteboards, Data · MCP Apps docs · @ClaudeDevs reset · TechRadar

Speech

ElevenLabs Scribe v3: 27% fewer word errors and real-time speaker labels

ElevenLabs released Scribe v3 for recorded audio and Scribe v3 Live for real time. v3 makes 27% fewer word errors than Scribe v2 across 102 languages; v3 Live cuts semantic errors by 43% versus Scribe v2 Realtime and adds speaker diarization (who said what) in real time for the first time.

  • Diarization: v3 makes 17% fewer speaker-labelling errors than v2 on recordings with three or more speakers.
    • On English real-world audio, v3 Live’s diarization error is 3.1% versus 5.9–7.3% for the other real-time models ElevenLabs tested.
  • Per language: German improved 42%, Spanish 35%, English 23%; on the real-time side Greek 77% and Hindi 36%. Polish isn’t singled out in the post.
  • Turn detection: v3 Live uses conversation context, not just silence, so a caller who pauses halfway through a number isn’t cut off.
    • Dutch energy firm Eneco saw postcodes recognised right first time 40% more often on real calls.
  • Controls: Up to 1,000 key terms per request, plain-English formatting rules (“write every date as YYYY-MM-DD”), digits or words for numbers, and a mid-sentence multilingual mode on Live.
  • Availability: In ElevenAPI as scribe_v3 and scribe_v3_live; EU, India and Singapore data residency and a zero-retention mode for enterprise.
    • ElevenLabs says it will open-source its 16-language real-world test set.

Why it matters: Speaker labels in real time were the missing piece for meeting bots and voice agents; without them live transcripts were a wall of text. All numbers are ElevenLabs’ own, scored its way (each language averaged equally), so compare carefully with other vendors’ tables.

Sources: ElevenLabs blog · @ElevenLabs

Energy

SpaceX signs Germany’s Singulus to build solar-cell lines for its Starmind satellites

Singulus Technologies announced on 10 November a supply contract with SpaceX for a turnkey heterojunction (HJT) solar cell production line, the first of several planned. The first line ships in the first half of 2027, and the partnership is meant to scale to more than 100 GW of production equipment for satellite manufacturing in Bastrop, Texas.

  • Who said what: SpaceX’s Starlink SVP Mike Nicolls said Singulus gives it “a high-performance manufacturing path for solar cells for our Starmind satellites”.
    • Elon Musk summarised it as “100 GW/year of HJT solar manufacturing equipment… for Starmind to create a sentient sun.”
  • Money behind it: The two have been partners since March 2026; SpaceX gave Singulus significant funding in March and again in October through a subordinated secured loan.
  • The tech: Singulus combines wet chemistry, vacuum coating and thermal processing in one platform, and pitches itself as a leading Western supplier of next-generation solar cell equipment.

Why it matters: “Starmind” is how SpaceX frames compute in orbit, and its hard limit is power. Buying the means to make its own cells at gigawatt scale, rather than buying panels, shows SpaceX is treating space-based AI compute as a manufacturing problem, not a slide.

Sources: Singulus press release · @SpaceX · @elonmusk

Autonomy

NVIDIA releases Alpamayo 2 Edge, an open 2B driving model that explains its decisions

NVIDIA published Alpamayo 2 Edge on 10 November, an open 2-billion-parameter vision-language-action model for self-driving that plans a driving path from surround cameras and can explain why, sized to run inside the car.

  • How it works: It uses what NVIDIA calls adaptive Chain-of-Causation reasoning plus a diffusion “action expert” that turns camera views, recent history and navigation input into trajectories, with explanations when needed.
  • Speed: NVIDIA reports 160–190 ms from pixels to trajectory on DRIVE AGX Thor after TensorRT optimisation.
  • Open weights: Published on Hugging Face as nvidia/Alpamayo2-Edge, with a technical blog on fitting reasoning models into onboard compute and latency limits.
  • Demo: NVIDIA’s video shows it yielding to pedestrians and making room for an oncoming bus while narrating its choices.

Why it matters: Reasoning driving models have mostly lived in the data centre. Getting one small and fast enough for in-car hardware, with open weights, gives smaller carmakers and researchers a starting point they don’t have to train from scratch.

Sources: Hugging Face model card · NVIDIA technical blog · @nvidia

Chips

NVIDIA tells cloud partners worldwide that 51% of new capacity must go to non-Chinese customers

NVIDIA has told cloud and data-centre partners globally that at least 51% of their services under future contracts must go to non-Chinese firms, the Financial Times reported. A B300 rack that sells for about $600,000 abroad reportedly went for $2.8 million inside China in October.

  • How it was delivered: Verbally, through NVIDIA’s sales and regional teams, not in writing. It applies globally, not only in Asia.
    • Existing leases are unaffected; only new deals must comply. NVIDIA hasn’t confirmed or denied it.
  • Other checks: NVIDIA has more than halved the number of Asian buyers authorised to purchase its AI chips, and Southeast Asian providers that want to become NVIDIA Cloud Partners now need its fleet-tracking software, Tom’s Hardware reports.
  • The squeeze: Chinese customers have reportedly ordered more than 6 million export-compliant RTX Pro 5500 cards, at least twice what NVIDIA could make in 2027.
    • A Singapore operator told the FT it’s “difficult to find enough non-Chinese clients outside the US.”
  • Why renting abroad works: Current US law gives regulators little authority over overseas cloud access, which is why Tencent, ByteDance, Alibaba and Moonshot have leaned on leases in Southeast Asia and Europe.

Why it matters: NVIDIA is doing privately what Washington hasn’t legislated, closing the rental route that has kept Chinese labs on frontier hardware. Neoclouds in Asia that depended on Chinese demand now have to find other customers or shrink.

Sources: Tom’s Hardware · Seoul Economic Daily · TechRadar

GitHub repos

  • elder-plinius/OBLITERATUS

    An open-source toolkit for finding and removing refusal behaviour from open-weight LLMs (“abliteration”). It probes hidden states to locate refusal directions with PCA, mean-difference, sparse-autoencoder and whitened-SVD methods, then projects them out, aiming to keep the model’s other abilities. It has a CLI (obliteratus obliterate <model> --method advanced), a Colab notebook and a Gradio Space, and optional telemetry that pools anonymous benchmark results. Pitched at red-teamers and alignment researchers; models it produces have their safeguards removed, so the responsibility for use is entirely yours.

    ★9.4k, AGPL-3.0, on today’s Python trending list

    View repo →
  • homeassistant-ai/ha-mcp

    An MCP server that lets Claude, ChatGPT and other assistants configure and automate Home Assistant. Since version 9.0 (28 October) it installs as an in-process integration from the new Home Assistant Marketplace (2026.11+), with optional local-only mode, Home Assistant sign-in or OIDC for remote access. Its own README recommends creating a dedicated admin account rather than using your personal token.

    ★6.8k, MIT, +790 stars in a day

    View repo →

Research papers

  • LingBot-Touch: heterogeneous tactile representation for dexterous manipulation

    One tactile foundation model that learns a shared representation across different touch sensors, producing plug-and-play “tactile tokens” that improve robot manipulation policies.

    2611.12806, 94 upvotes

    Read paper →
  • Sage-Taiyi: a metacognitive architecture for scientific discovery

    A system that monitors its own accumulated understanding to decide what to investigate next and how.

    2611.12762

    Read paper →
  • Flide: exact state forking for decision-model agents

    An inference engine that speeds up LLM agents exploring many branches by separating the logical request tree from physical execution, with “ForkAttention” keeping each branch’s results exact.

    2610.42148

    Read paper →
  • Infrathon: can agents optimise real AI infrastructure over long horizons?

    A benchmark of long-horizon infrastructure optimisation tasks for frontier agents.

    2611.11995

    Read paper →
  • InfraRSI: recursive self-improvement for training infrastructure

    Applies self-improvement loops to distributed training, where launch settings, kernels and communication schedules have to be tuned together rather than one at a time.

    2611.11558

    Read paper →

AI workflows

1. Seven Claude Code routines keep a one-person app healthy · Workflow page

  • Replaces: evening upkeep of a solo product: checking search traffic, security, production bugs and data freshness by hand, then writing the fixes.
    • Last week 46% of the routines’ pull requests merged on proof alone, and the rest took about 3–4 hours of review.
  • What to connect: Claude Code routines (scheduled cloud sessions) on a repo with real tests; Search Console (mcp-gsc), PostHog error tracking, and Supabase and Cloudflare APIs (read-only tokens added by the network proxy).
    • Also: GitHub Actions, and a shared “notebook” branch for state.
    • Only machine-verifiable changes auto-merge, a weekly Slack digest replaces daily alerts, and a /decide skill turns long parked issues into one-click questions.
    • Cost: Claude Max ($200/month), about 200 agent minutes a week, ~5,000 Actions minutes a month.
  • Who it’s for: solo founders and small teams running a live web product who want fixes, not alerts.

2. A bot drafts the invoices, a person still presses Send · Workflow page

  • Replaces: monthly billing for a tutoring side business: reading the calendar, working out which sessions are billable, typing invoices and updating the books.
  • What to connect: a scheduled Grok Bot routine, the Google Calendar connector (read-only), a PayPal developer app used only to list invoices and create unsent drafts, and a Python script that writes a Quicken QIF import.
    • Calendar events are tagged (cancel, no-show, time away, cash) instead of deleted, existing invoices are checked to avoid double billing, and the owner’s weekly lesson count beats the bot’s count when they disagree.
  • Who it’s for: sole traders and service businesses that bill per session or per job.

3. One AI “webmaster” instead of a team of agents for a winery’s site and shop · Workflow page

  • Replaces: running a family winery’s website and WooCommerce shop with several agents (SEO, store, content) that lost DNS changes and the blog calendar in handoffs and burned the whole usage quota in 2 days.
  • What to connect: one Grok Bot with a job description and a one-sentence mission; GitHub plus a cloud coding agent, Vercel, the WooCommerce REST API, Search Console, Bing Webmaster Tools, GA4, Google Ads and Merchant Center.
    • Notion as memory (SEO plan, blog calendar, product-page database, change log).
    • Fixed weekly routines (10 product pages a day, SEO Monday, blog Tuesday/Friday, Ads Wednesday), with prices, stock, spend and DNS needing the owner’s OK, and no emails or payments ever.
  • Who it’s for: small online shops and family businesses that want one agent to run their site and store.

Quick hits

  • Grok Bot joins your tailnet — Grok Bot can now join personal or company Tailscale networks, so it can reach private MCP servers, internal tools and self-hosted apps without exposing them to the internet. @bot
  • OpenAI’s cyber keynote is Thursday — OpenAI says it’s turning lessons from defending itself into customer tools, to be announced at a live “Cyber at the Frontier” keynote on 12 November. @OpenAI
  • Llama for Windows on the .NET Conf stage — @jeffboudier announced Llama for Windows at Microsoft’s .NET Conf keynote: a free, open-source Windows app built on llama.cpp that downloads and runs local models with no terminal, opened with Alt + Space.
    • The app itself was first published in August. HF blog
  • Hugging Face passes 20M builders and 300 PB — Clem Delangue says the Hub now hosts 300 PB of models, datasets, agent traces and RL environments (9 PB added in the last week).
    • October was its biggest month for users and revenue.
    • The top trending model is Idol, a 600M-parameter image-to-3D model from Spain’s Sperid Labs that needs under 7.2 GB of VRAM and makes 1M splats in 3 seconds, Apache 2.0. @ClementDelangue · @NielsRogge
  • Unreal Engine 5.9 preview runs in the browser — Tim Sweeney says the preview brings AI acceleration work ported from the unfinished Unreal Engine 6 plus a new web export.
  • Factory’s Security Factory — Factory opened a private preview that links security investigation, fix and verification into one workflow, turning findings into tested patches.
    • @dexhorthy framed it as a race against attackers “armed with abliterated model and sophisticated attack tooling”. @factory via @dexhorthy

Go deeper

Get the free daily brief

↑