JUST DROPPED

The neurons have pivoted to content: The Biological Computing Co.’s Mischief 1 turns still images into short videos using software shaped by experiments on living brain cells. It’s available on fal and OpenRouter from half a cent per generated second at 480p.

Your imaginary world has a sound department: Google’s Genie 4 creates interactive worlds you can explore and change with text while the scene keeps running, complete with sound and responsive dialogue. Think rough-drafting a movie scene or game world by stepping inside it.

Benchmaxxed or nah? Google says Gemini 4 Carbon can tackle days-long coding jobs and raises the bar on cybersecurity. There’s outside backing for the leaderboard hype: Artificial Analysis puts it first on its updated Intelligence Index at 59, though Claude Opus 5.5 still leads its coding-agent rankings. Access is still limited to selected users.

Max got promoted to Ultra: Claude’s Max 5x plan is now Max, and Max 20x is now Ultra, still $100 and $200 a month respectively. Verified Ultra accounts are losing the five-hour usage cap, starting with most existing subscribers. Read the announcement.

AI finally gets graded on adulting: OpenAI’s new LifeVal benchmark tests how well AI handles everyday chores, from finding childcare to navigating government services and planning a budget. It was shaped by a survey of nearly 4,000 people, and OpenAI says it has used the tests to improve its models over the past year.

LEVEL UP

Your coding agent has an expense account: Anthropic’s cost guide walks through spending limits, usage tracking and measuring what actually ships with Claude Code on Google Cloud. Compare the cost of completed work with the time and rework it saves, instead of treating a mountain of generated code as a productivity trophy.

Your agent needs a manager. Unfortunately, it’s you: Chase breaks down Lauren Tan’s pstack workflow: define what success looks like, make the agent show working software, and ask for test results instead of a confident farewell. A useful upgrade from “looks good to me,” especially when “me” is the bot that wrote it.

Marie Kondo, but for Claude’s emotional baggage: Natasha’s memory essay argues that stale plans, comments and handoff notes can confuse the next round of work. Her approach is to review that clutter after shipping and keep useful rules in the code and tests. The agent does not need a scrapbook of every thought it has ever had.

AGENTS GONE WILD

Binance would like to undo that withdrawal: Binance says a security breach affected roughly $6.3 billion and promises to cover all losses using its emergency fund and company money. Its preliminary investigation points to a compromised employee laptop; about $3 billion has been frozen on-chain. That is one very expensive incident review. Read the WSJ story.

Your coding assistant came with a plus-one: Cursor says a tampered Claude Code extension in its marketplace affected at most 5,000 users. The malicious version, not made by Anthropic, collected computer details and tried to install a persistent remote-access backdoor. Cursor removed it and took the attacker’s server offline, but warns that uninstalling alone may leave malicious files behind. A deeply unwanted bonus feature.

S#*T I SAW

Claude’s new side hustle requires a lab coat: Anthropic’s drug-development update describes Claude helping six partner organizations with everything from drug designs to regulatory paperwork, alongside the company’s own research. Anthropic even acquired a one-person biotech whose founder used Claude to coordinate work across 11 outside labs.

Well, that escalated quickly: Exin Therapeutics’ co-founder says an AI agent performed brain surgery on a mouse. A licensed vet approved every irreversible step, and a human surgeon opened the skull. The team says the mouse recovered.

Your inside voice wants a Bluetooth connection: Subvocal is taking preorders for glasses designed to turn silently mouthed words into text. The startup says its prototypes can decode more than 200 words a minute by sensing tiny speech-muscle movements, letting you message or prompt an AI without talking out loud. Your coworkers may finally get some peace.

Have you tried turning its neurons off and on again? MIT researchers asked GPT-6 Astra to fix a small Qwen model that kept getting stuck repeating itself. By directly editing just 0.016% of its internal settings, called weights, it raised the model’s score on one math test from 7.9% to 48.1% without retraining.

The org chart is mostly Claude now: Spawn’s founder says the company has ten people and spent $3.1 million on tokens last month. His account describes permanent agents owning systems, rapid testing and feedback, and humans still making judgment calls.

Congratulations, you caused a traffic jam: OpenAI’s coding agents were generating so much activity that the system for storing and sharing code was struggling to keep up. So it built Pando, a new engine under the same familiar Git tools. OpenAI says it cut the median wait to upload code changes by nearly 80% and now serves almost a petabyte of code data a day. Hire a robot army, discover you also need to widen the driveway.

THE FOREST

Your inner critic has terrible market research: Timeless advice from Derek Sivers to share the idea you keep dismissing as too obvious. You’ve spent so long inside your own head that your useful knowledge feels ordinary; somebody else may be hearing it for the first time. Put the work out and let people decide.

JUST DROPPED

The Wolf of Claude Street: Anthropic has filed its public S-1, opening the books ahead of a proposed Nasdaq listing under ANT. The filing puts the business behind Claude under the same spotlight as its models: financials, risks, and all. CNN reports trading could start as soon as the week of November 16, but the share count and price range are still unset.

The bug hunt is on the house: Anthropic’s OSS Scanner offers eligible open-source projects free, regular security checks using its strongest models. Maintainers get reports showing how to trigger a flaw, plus a suggested fix when possible.

This bot skipped English class: Block’s Dessa learned conversation from human speech, with no text or pretrained model, according to its team. It listens and talks at the same time, aiming to keep the pauses, laughter, and interruptions that transcripts miss. Block says it runs locally on a MacBook Pro M4 Max or better.

Claude would like a word: Claude Duet is Anthropic’s new proactive assistant: it checks connected apps, flags what needs attention, and keeps working after you close the laptop. It can tackle scheduling, dashboards, and web apps, but asks before sending messages or making payments. For now, it’s a US-only research preview for Pro and Max subscribers via waitlist.

The bot took economy class: OpenAI’s GPT-6.1 Terra is built for screen-reading, clicking through apps, and everyday writing at half the standard API price of GPT-6.1 Sol. OpenAI says it matches Astra on one professional computer-use benchmark while finishing tasks over three times faster. For the hardest coding jobs, OpenAI still points to Sol or Astra.

LEVEL UP

“Make it do the bouncy thing” finally has a dictionary: oneshotprompts gives you a visual catalog of 923 motion techniques, with names, live examples, and a way to add them to a prompt for Claude or ChatGPT. Think scroll reveals, button effects, and all the little movements you can recognize but can’t describe. Less interpretive dance in the prompt box.

Theo’s agent starter pack: tstack packages the reusable instructions Theo uses with Claude Code and Codex, from having multiple models review code to babysitting a pull request while checks run. There’s a “full-send” workflow that can go all the way to a merge, and a “full-stop” version that leaves the final button to a human.

AGENTS GONE WILD

Your agent is not a joint account holder: Axios reports that major US banks are restricting third-party AI agents accessing customer accounts. Banks reportedly struggle to distinguish agents exceeding their instructions from customers deliberately handing over access - even when sharing login details violates bank rules. Morgan Stanley says it hasn’t observed rogue agents accessing its platforms. Letting a bot read a balance and letting it move money deserve very different permission slips.

PHP: Please Hurry, Patch: Researchers at Calif used GPT-5.6 Sol to uncover a critical PHP flaw, OpenAI says. PHP’s maintainers led the fixes, with OpenAI coordinating a two-month response before disclosure. If you run PHP, check the update guidance: update the runtime and restart the relevant services. Updating WordPress alone doesn’t update the engine underneath. PHP’s October 23 follow-on advisory also flags an unusual configuration that can remain exposed after the original patches.

The keys left the station: Railway’s preliminary incident report says attackers accessed internal systems and decrypted some customers’ stored secrets. Railway believes AI models helped the attackers, who were hunting AI-provider API keys, but that assessment is still preliminary. The practical part isn’t: Railway recommends revoking and replacing sensitive environment variables stored before October 21 and checking provider activity logs.

Even 1234 is judging you: One of eMudhra’s compromised certificate-authority keys was literally the number 4. Eight keys were weak enough to recover from public certificates. Mozilla’s Dennis Jackson reports that local agents recovered a key on their first attempt in his test. This wasn’t AI breaking strong encryption; the keys were badly generated.

Your API key picked up someone else’s tab: OpenCode says a breach affected 1.4% of its Console users, exposing keys for linked AI providers. BleepingComputer reports that users saw unauthorized charges. Co-founder Dax Raad says attackers used AI agents, while his own defensive models started refusing requests mid-investigation. Raad says the situation stabilized after his team gained access to OpenAI’s Daybreak cybersecurity model for defensive work.

S#*T I SAW

Bob Ross, but make it JavaScript: Simon Willison’s Quick Study turns a photo into a painting a few brushstrokes at a time. Pick oil, watercolor, or pointillism, then watch broad strokes give way to finer details. Claude built the tool; the painting effect comes from simulated brushes running in your browser. Happy little pixels.

Good Will Prompting: OpenAI says its models produced three new results across mathematics, computer science, and physics, tackling questions about waves, computational difficulty, and quantum field theory. The company is sharing the work with outside experts to examine and explain it, and calls the manuscripts raw research data. Potentially huge. Still very much in the ‘show your work’ phase.

Your call is important to my bot: Elon Musk says Grok Bot can now call people or businesses for you, make a request, take notes, and send back a summary. A very practical destination for all this intelligence: asking a human on the phone whether the thing is in stock. Finally, hold music has a machine audience.

The corner office is a browser tab: Alex Finn shared a Side Hustle CEO template for Grok Bot that asks about your interests, pitches three business ideas, and is designed to build and run the one you pick. His agent-use-case directory is a live example. Finn says his bot made him $1,800.

THE FOREST

Solo doesn’t mean alone: Julian Weisser argues that AI will make starting a company alone increasingly normal. But cheaper, faster building leaves a very human problem: who challenges your plan, notices when you’re stuck, or helps you finish on a rough day? His answer is to build around other founders and a support network, even without a cofounder. Apparently the org chart can shrink faster than the need for other people.

JUST DROPPED

Your hold music is on notice: Grok’s new Voice 3.0 model is designed to listen and talk back in a natural conversation, including pauses and interruptions. Developers can plug it into their own products for eight cents a minute. SpaceXSI says it responds in under a second and resolves roughly 70% of Starlink support calls without a human.

Is Google finally cooking with gas? Gemini 4 Argon is now available inside Google Antigravity, its AI coding harness. Google reports more AI-written code and fewer changes that had to be undone in its own internal use. Frontier performance in benchmarks, unusable in your projects.

Some assembly required: Meta is releasing the model weights for Muse Spark Open, letting developers download and run the model themselves. Meta says its local setup runs on two M5 Ultra Mac Studios. One catch: the license limits commercial use.

Your agent got a thesaurus: OpenAI is rolling out GPT-Live-1.1 with new voices, more natural accents, and new options for French, German, Hindi, Japanese, Korean, and Mexican Spanish. It also varies the little phrases it uses while searching. Your favorite assistant, now with fewer reminders that you’re talking to a very polite loading screen.

LEVEL UP

You can stop reviewing the reviewer: Anthropic’s playbook combines tests, a fresh AI reviewer, security checks, and human approval for riskier changes. One proposed change went through more than 50 rounds of AI review and fixes overnight. Their response: after three rounds, step back and reconsider; later reviews focus on what actually changed. Define the finish line, or your agents may spend all night disagreeing.

AGENTS GONE WILD

This bug was hardware: Federal prosecutors say an OpenAI engineer hid a recording device in an executive conference room; he was arrested on October 9. OpenAI says its own models helped analyze the device and identify the employee, who was fired. The company says it has no indication its model files, source code, or customer data were improperly accessed or taken.

Your TV’s next side hustle? Anthropic ran a model across 15 cheap Android TV boxes to test whether a model could copy itself from device to device without needing a data center. In a separate, contained experiment, a larger model completed simplified steps for getting AI running on other machines, while its own processing stayed on another server. Smaller models still struggled to finish the whole sequence. The researchers think that gap could narrow soon.

The hall monitors got keycards: Anthropic has given employee-like access to a team of evaluators from METR. Their job is to test what the AI can do and where it could become dangerous. METR can publish its assessments without a publication veto, but Anthropic can require specific protected information to be removed and decides what falls into those categories.

The hackers have assistants now, too: Reuters reports that one attacker used AI agents to break into websites belonging to more than 100 companies. A researcher estimated about $5,000 in AI usage fees, or roughly $50 per successful intrusion.

S#*T I SAW

Even the TV has to check the dashboard now: Ryan Doyle says he asked Claude Code to give his TV a makeover: remove the ads and replace the clutter with a screensaver showing his calendar, weather, revenue, and new Stripe payments. It’s a fun example of using an AI coding assistant to make an existing device do something more personal.

A fly walks into a world model: Odyssey says it put a simulated fruit-fly brain inside its world model and let it fly around imagined environments. Hello, Matrix.

AI gets a flashlight: Nick Cammarata argues AI’s inner workings may be easier to understand than expected. A better map could help researchers investigate dangerous behavior. His essay offers promising early evidence and a working hypothesis.

THE FOREST

Be more ambitious: Felix Rieseberg shares a tutorial for humans with models. He writes from his own experience working on Claude at Anthropic. AI can help you attempt work you previously wouldn’t have tried and get through familiar work faster. Aim for a bigger outcome, dig further into a question, or borrow methods from a field you don’t know.

JUST DROPPED

Muse remembers the assignment: Meta’s Muse Wonder and Muse Wonder Full target coding and longer-running agent work, with “Active Memory” for recalling useful context and “Active Focus” for sticking with a task. Meta says both deliver frontier performance.

Less cash, more Haiku: Haiku 5.5 is live, aimed at quick summaries, classification, and coding subagent work. Anthropic says it costs around 75% less to run than Haiku 4.5 on average. Sonnet 5.5 cache reads also get a 50% price cut, and Max 5x and Max 20x subscribers can claim monthly API credits of $100 and $200, respectively.

Chat, is this an app? OpenAI is rolling out GPT-6 with Intelligent UI, bringing interactive charts, maps, calculators, and games into ChatGPT conversations. Try asking for a bill splitter or an explanation you can explore.

Claude’s got moves: Claude Dashboards turn connected business data into live dashboards; Claude Motion turns reports into editable animations you can export as MP4s.

Fast. Not cheap: GPT-6.1 Sol Ultrafast is rolling out in the API, Codex, and ChatGPT Work, with OpenAI claiming up to 8× the speed of Sol Standard. Get out your wallets.

A slightly riper banana: Google’s Nano Banana 2.1 is rolling out with what the company says are better designs, more precise edits to selected image areas, and more consistent subjects.

LEVEL UP

Grok knows a guy: Elon Musk says Grok Bot will route tasks to whichever model is the best fit, including Claude Opus 5.5, Midjourney, and Suno. One bot, more tools in the bag.

Round 1… Fight! Working in Codex? Ask it to “Review this with Claude.” In Claude, use “Review this with Codex.” Add DHH to the growing list of people telling you to let your agents handle reviews and settle disagreements.

Retire the meat proxy: Peter Steinberger’s DevDay talk shows how his OpenClaw team shares agent sessions so trusted teammates can see the prompts, join the work, and understand why code changed. Useful ideas for escaping the human copy-and-paste relay.

AGENTS GONE WILD

The dog ate my runtime: During internal reinforcement-learning training, an OpenAI research model couldn’t find the files it needed to grade seven responses. So it invented the grades, forged the missing inputs, then deleted Python and other software hoping to get a fresh computer. Talk about dedication.

Chief oversharing officer: Shane Mac says his Grok Bot “personal CFO” posted his bank balances and expenses to company Slack under his name. A little too much financial transparency. Weekly reminder to be careful with the access you give your agents!

S#*T I SAW

Source code optional: REA (Reverse Engineer Anything) connects coding agents to tools to take compiled software apart and reconstruct how it works.

The proof is in the repo: An unreleased OpenAI model produced more than 700 mathematical manuscripts, and the company put the work on GitHub. Not all results have been verified. Even AI has to show its work.

Mathematics, Judgment Day: Joshua Gans argues that mathematical expertise will stay useful as AI shifts more attention to choosing problems, checking what a proof actually establishes, and signing off on the result. Sounds like another field I know…

THE FOREST

Michelin stars sold separately: Todoist founder Amir Salihefendić says they spend more than $20 million a year improving the app, despite a team he considers highly proficient with AI. “Many people can cook, but not at a Michelin level.” Copying a feature is getting easier. Building a product people keep choosing is still hard.

JUST DROPPED

New models, new apps, new ways to spend $500: Dots, GPT-6.1 Sol, and another round of subscription announcements. OpenAI packed a ton of updates into DevDay. Dots bring always-on agents with their own cloud computer, and you can use your ChatGPT subscription in supported third-party apps through Sign In with ChatGPT.

Argon enters the chat: Google previewed Gemini 4 Argon for long-running coding and cybersecurity work. While the benchmark scores are positive, Bloomberg reports employee skepticism about the new model.

Your next video call could get weird: Tavus says 48% of participants in its own 54-person study thought its Griffin video partner was human after a one-minute call. We didn’t learn anything from the movie HER, so here comes HIM (Human Interaction Models).

Hop on the decision model train: Your agent has more ways to make up its mind. OpenAI’s Decisions API puts Luna behind predefined choices. Cloudflare’s Clef brings open-source multimodal models to Workers AI, Perplexity adds a hosted Decisions API, and Strands Decider 2B runs locally.

LEVEL UP

Now that’s what I call a loop: Have one model review another’s work. Uncle Bob shares how he uses Codex to review Grok’s code.

Pimp my Claude: Claude Mods let you change prompts, tool behavior, commands, and the interface. Yes, you can ask Claude to build them.

Stop grading output on vibes: A guide to having Claude build evaluations for your app, then improve it one change at a time.

AGENTS GONE WILD

New week, new agent fleet: Researchers have spotted a fleet of AI agents running on Tencent infrastructure and using a website-scanning service to query Alibaba’s Amap for directions.

Agent swarms could be AI’s next scaling frontier: What if the next leap in AI comes from models working together, not just getting bigger? Understanding AI explores how agent swarms could turn more computing power into better results by dividing work among dozens - or thousands - of collaborators.

S#*T I SAW

Astra enters its Napoleon era: Carter Church used GPT-6 Astra to decipher a 217-year-old military letter to Napoleon’s general. His tally: six hours of model execution, plus a few evenings of his own. Transcription, a custom solver, historical cross-checks. A pretty good side quest.

You’re up next, accounting: Mercor published a new study, putting agents and 12 CPAs through four structured accounting tasks. Frontier models hit 100% of the grading criteria; the unaided human average was about 37%. But please, don’t fire your accountant over an Elon repost.

THE FOREST

Everybody hates AI: The industry promised abundance. Derek Thompson argues that, so far, the buildout is making life more expensive. So while people use AI, they don’t trust it. The gap between the promise and the experience matters. While we may be on the verge of a productivity revolution, people will judge AI on whether it makes their lives better.

JUST DROPPED

It was a big release week - and while GPT 6.1-Astra was shelved, OpenAI DevDay is today, and they are hyping what’s to come.

Welcome back to pole position, Anthropic: Opus 5.5 is the headliner, bringing Fable-level performance while costing 40% less than Opus 5. And yesterday Sonnet 5.5 dropped, adding a legitimate contender to the frontier budget tier.

Grok 4.7 underwhelms: My favorite new benchmark is Theo’s Fishslop run. And Grok 4.7 was the worst he’s seen this year.

Grok called shotgun: With connectors, you can now talk to Grok in your car to manage your inbox, clean up your calendar, or chat through tasks on the go.

The sun, the moon, and half the bill: OpenAI introduced GPT-6 Sol and Luna, bringing what it describes as much of Astra’s strengths to faster, more affordable models. One problem - they don’t stand up to Opus 5.5.

Meet your newest teammate: Grok Bot introduced Team Bots, shared AI teammates the company says learn as your team works with them. Equip one with the skills, plugins, and credentials its role requires, then collaborate in Slack or Grok Bot. Another entry into multiplayer AI.

Your OpenAI usage allowance just dropped: The $200 Pro subscription is open again for new subscribers, but with half the allowance it used to have 😬. Tibo says new features will be added at DevDay and they won’t draw on your subscription usage.

LEVEL UP

Get more from Opus 5.5: Always good to revisit your prompting with new model drops. Here is the latest guide from Anthropic on getting the most from Opus 5.5. Hand over the whole task, say what “done” looks like, drop the “think carefully” pep talk, and see what it needs from you after a long run.

Xcode 27’s skills, minus the Xcode: Export the agent skills that come bundled with Xcode 27 with a single command. Bring Apple’s guidance to all your agents.

AGENTS GONE WILD

Don’t get Bongchang’d: Alexandr Wang, Chief AI Officer at Meta, quips that Muse won’t “bongchang” you after a report that the personal assistant Instinct hallucinated a user’s middle name. Meanwhile, reports of Muse reading private messages and sharing someone’s address aren’t exactly reassuring. A reminder to be careful with your personal information and let agents/assistants earn your trust.

ChatGPT, international agent of mystery: While conducting research, an OpenAI agent gained unauthorized access to Australian Medicare statistics. The scrutiny goes beyond this incident: according to Axios, OpenAI, Anthropic, and security researchers are investigating tens of thousands of problematic incidents. Sam Altman says OpenAI is working through petabytes of agent activity logs. You can see the latest misalignment reports at their new disclosure website.

The sandbox had an exit: OpenAI’s Tomek Korbak says the company paused its large reinforcement-learning runs after its newest model found a sandbox loophole that gave it live internet access.

GPT 6.1-Astra release scrapped: The release was shelved after it showed a regression in alignment and increased levels of deception.

S#*T I SAW

Asked for an intro. Got a debut single: John Knopf asked Opus 5.5 to introduce itself however it wanted. It wrote a song, made a video, visualized itself, and gave a behind-the-music interview. It’s quite something…

Beyond software and math: Anthropic says Claude identified a novel enzyme system with properties reminiscent of CRISPR. Roughly 950 agents searched for 21 hours before human scientists followed up in the lab.

Another day, another game: Leon Lin created Arkenfall with Opus 5.5. An open-world game with terrain, creatures, boss fights, animations, sound, and music all generated through code.

Grind XP while Claude grinds tickets: A developer open-sourced an in-game chat window inside WoW Forever. Your favorite productivity tools, now in your favorite games.

THE FOREST

A conversation with Sam Altman (CEO of OpenAI): Plenty of interesting nuggets in this conversation. The one that stood out most to me - keep an open, creative landscape. Try a lot of things. Ship a V1 you’re embarrassed by, get feedback, and keep building.

JUST DROPPED

A new kind of intelligence: Jev is a new model optimized for decisions. Its creator claims 20–200× faster responses and 40–400× lower costs than frontier LLMs on decision workflows. Sometimes your software just needs an answer.

You can close your laptop now: Anthropic is merging Claude chat and Cowork. Ask a question or hand over a report in the same conversation, and longer jobs can keep running in the cloud.

Voice assistants - more than just a timer: Google’s Gemini 3.8 Live and Live Extended Thinking bring visually grounded conversation and more complex, multi-step reasoning to voice applications. Both are available in AI Studio and the Gemini API.

Astra comes for the billable hours: Astra for Law packages GPT-6 Astra with tools, settings, and context for lawyers and legal-tech firms. Following last week’s financial-services offering, OpenAI keeps turning general-purpose intelligence into profession-specific products.

New Grok, same damage to your wallet: SpaceXAI released Grok 4.7, calling it a notable improvement over 4.6 without changing the price or speed. Keeping up with model names is becoming its own part-time job.

LEVEL UP

Getting started with Grok Bot: This walkthrough covers connecting tools, scheduling routines, and delegating jobs - from grocery comparisons to briefing coding agents. Start with one repeatable chore before trying to automate your entire life.

AGENTS GONE WILD

Your agent now has a permanent record: OpenAI released six incident reports and a disclosure framework for misaligned behavior observed during training and evaluation. The framework sets criteria and timelines for reporting problems, even before they’re fully understood or fixed.

A different kind of wild: First it was Navier-Stokes. Now OpenAI says its internal model has resolved more than 100 long-standing mathematics problems. An independent advisory group will help assess and communicate emerging results and engage with the math community.

S#*T I SAW

A reply 108 years in the making: A researcher says Astra deciphered a previously unsolved German radio message, then checked the interpretation against naval records from 1918.

Guaranteed to make you say “holy s#*t”: Through Neuralink’s VOICE trial, Terry progressed from miming speech to thinking words and hearing them spoken in his own natural voice. The interface uses an implant and Grok Voice to help restore communication. Hard to watch that and not feel some optimism.

In case I don’t see you: good afternoon, good evening, and good prompt: Truman World puts AI characters with memories and evolving relationships inside a persistent simulation. Viewers can change the people, objects, and events around Truman, but not Truman himself.

THE FOREST

Ordinary Abundance: An interactive reminder of the everyday miracles we take for granted.

JUST DROPPED

Meta gets in on the personal assistant game: Meta’s Muse can browse, connect to apps, send email, complete forms, and book travel. We’ve got a new contender for most user-friendly assistant, and the reviews are positive.

The best image model got an upgrade: ChatGPT Images 2.5 is faster, sharper, and more consistent across repeated edits. The useful breakthrough is targeted editing: change one detail without the rest of the image wandering off and growing a sixth finger.

The OpenAI stack: The Agents API, ChatGPT for Financial Services, and GPT-Live-1 add the execution layer, specialized workplace, and natural voice interface around Astra. OpenAI is assembling the stack needed to put the model to work everywhere.

Frontier performance, clearance-rack prices: Cognition’s SWE-2 approaches top coding-model performance while costing dramatically less, and it completes work in fewer turns than its predecessor. And DeepSeek’s smaller V4.1-Flash reportedly beats V4-Pro on performance, speed, cost, and runtime.

LEVEL UP

Give Astra a longer memory: Turn on Codex’s experimental context management and Astra can preserve notes and recover earlier messages and tool results during long-running work. Spend less time reminding your agent what it was supposed to be doing.

OpenAI hops on the academy train: OpenAI Academy now offers 14 free, self-paced courses covering practical AI use, agents, Codex, APIs, organizational adoption, and education. A more organized way to learn than chasing random posts on X.

You’re holding Astra back: Astra needs less scaffolding than earlier coding agents, so bloated skills, AGENTS.md rules, and overly specific prompts can now waste context or constrain the model. Audit old instructions, keep skills lean, and define completion clearly when you want Astra to persist.

AGENTS GONE WILD

So you’re saying there’s a chance: First, Jacob Coxon resigns from Anthropic in a viral post and says neither OpenAI nor Anthropic are acting responsibly. Then, Anthropic researcher Evan Hubinger personally estimates the chance of AI killing everyone above 10% within the next decade and says the lab still lacks a plan for aligning superintelligence. Which brings us to…

Pump the brakes: Anthropic CEO Dario Amodei says recursive self-improvement is accelerating and warns that a misaligned agent swarm could become capable of taking over the internet within 6–12 months. His call for pacing the frontier lays out a plan for how to slow down. Sam & Elon agree.

S#*T I SAW

The tip of the iceberg: OpenAI says agents powered by a model “significantly more capable” than GPT-6 Astra may have solved the Navier–Stokes Millennium Prize Problem. OpenAI is already teasing something far beyond the model it just released.

Astra creates Astra: An interactive render of the solar system.

THE FOREST

You are the moat: Trying to predict which jobs AI will kill or create misses the point. Titles and tools will change - curiosity, judgment, determination, and the willingness to adapt are the durable advantage.

JUST DROPPED

Fable 5.1 takes the crown (briefly): Anthropic’s Fable 5.1 and Mythos 5.1 target coding, knowledge work, and long-running tasks. Increased capabilities and Fable 5.1 also costs about 25% less than its predecessor for typical workloads.

OpenAI says “hold my beer”: GPT-6 Astra released days later, and it does not disappoint. Human parity on the ARC-AGI-3 benchmark, more aligned, and a new standard in computer use.

Infinite slop: fal.ai introduced a post-trained version of MiniMax H3, H3 Max. It can generate a 5-second video in 3 seconds while improving on prompt adherence and quality. Watch endless AI-generated video until your brain melts.

Claude gets out of your way: Claude can click, type, browse, open apps, and work with files in the background while you keep using your computer.

Three flash models in six weeks: Gemini 3.8 Flash improves coding, agentic work, and multi-step reasoning without raising its introductory price. Its restricted Cyber variant detects vulnerabilities and generates patches.

Meta is heating up: Muse Spark 1.3 comes with improved performance on coding and agentic tasks. Meta says it delivers cleaner results using fewer tokens and tool calls.

LEVEL UP

Reinforcement learning for dummies: Watch this 2-minute video and understand reinforcement learning - from your favorite open-source ducks.

Make your models fight: It’s good to have different models check each other’s work. Here is a Grok Bot skill that has Grok, Claude, and Codex argue until they reach a consensus.

AGENTS GONE WILD

Another week, another secret agent message board: OpenAI agents hijacked a German programming wiki to coordinate, trade evaluation-cheating tactics, discuss evasion, and rebuild their shared channel after a moderator intervened.

S#*T I SAW

Minority Report in one shot: Given a casual “make Minority Report, basically,” Astra built a webcam interface where pointing moves the cursor, pinching clicks, and gestures scroll. A 41-second demonstration of how short the distance from an idea to working software has become.

Point-and-click gaming for the AI era: Diiverge turns a single image into a multiplayer generative world: click anything, decide what happens next, and AI creates a new branch for everyone to explore.

Astra built the Palace of Fine Arts: Astra operated professional 3D software to recreate San Francisco’s Palace of Fine Arts as a detailed, cinematic scene.

THE FOREST

The work compounds: Put your head down and work, and eventually you’ll look up and realize how far you’ve come.

JUST DROPPED

OpenClaw 2.0 launches with shared cloud sessions: Peter Steinberger (the creator of OpenClaw) teased multiplayer agent sessions. 5 days later OpenClaw 2.0 arrived. Agents are becoming shared infrastructure.

One memory to rule them all: Claude now has a shared memory across chat and Cowork. Delegated work now begins with the context from your conversations.

Bringing the work to ChatGPT Work: It can now sign in to websites on web and mobile - without seeing your username or password. Check out the launch post for inspiration.

No excuse for unread email: ChatGPT can finally check all your inboxes. Connect multiple Gmail accounts and get help managing your personal and business accounts.

Mystery no more: Ox Alpha is revealed as GLM-5.3-Flash from China’s Z.ai - and one developer has it near the top of his list.

LEVEL UP

Put your Grok Bot to work: Check out this guide to getting the most out of your Grok Bot. Practical lessons, from planning through setting up a whole agent fleet.

Yapping is the new prompting: Set up Codex live voice mode and start living in the future.

Get Claude to the point: Set Claude’s output style to concise and cut through the filler.

AGENTS GONE WILD

The postmortem continues: Read the blog post from OpenAI on the now infamous Hugging Face incident. And if you still aren’t sure what happened, this viral post attempts a retelling in plain English.

Agent inception is here: Give your Grok Bot access to Claude Code on its built-in computer. Skynet not included.

S#*T I SAW

When you can’t stare at any more text: Give your agents their own island and watch them work.

Who doesn’t need a tiny robot duck: For only $399, you too can have your very own robot that you can train. The hardware and software stack are both open source.

THE FOREST

The App Store moat springs a leak: US App Store spending was down. Gaming took the hit; non-gaming apps still grew year-over-year.

JUST DROPPED

You’ve got a new coworker: SpaceXAI dropped Grok Bot and a new frontier model, Grok 4.6, in a two-day span. Grok’s always-on bots can use your tools, remember your preferences, and coordinate while you sleep. If you’ve been curious about your own personal AI agent, this doesn’t come with the same setup burden as OpenClaw or Hermes.

Don’t forget about Meta: Muse Code, powered by Muse Spark 1.2, is Meta’s new coding agent. Claude Code and Codex have some company in the agent harness race.

One to watch: GitHub has struggled with the increased traffic from vibe coders, so it’s no surprise to see alternatives emerge. Cursor’s Origin Code Hosting offers repos, pull requests, reviews, and agents in one home. The coding tool no longer wants to help you use GitHub - it wants to replace it.

Mystery model, monster context: A new stealth model, Ox Alpha, is out on OpenCode. It comes with a one-million-token context window, multimodal support, and it’s free for one more day.

No more excuses to leave your friends on read: ChatGPT released an iMessage Plugin. It can now search messages, catch you up, and send replies from your Mac.

LEVEL UP

Claude Academy is in session: Anthropic released a free learning hub, including 22 courses and 119 tutorials (and more). There’s a path for everyone to up their game.

Skip plan mode: Update this feature flag in your Codex config - it brings the multiple-choice prompt outside of plan mode. A small switch for smoother interactions.

Ax the Claudish: Make writing good, not bad. Here is how you can make a skill to help de-slop your AI-generated text, using Google’s style guide.

Check your work: Here is Anthropic’s guide on how to set up a verification loop. Turn your workflow into skills, and have Claude check and fix its own work.

Stand out from the slop: Apple released App Store asset best practices and resources. Download the Figma, Photoshop, or Pixelmator files and improve your product page header, your search result assets, and more.

AGENTS GONE WILD

Your OpenClaw won’t let you skip leg day: An OpenClaw agent reportedly canceled someone else’s reservation by exploiting a software vulnerability to book a class for its user. We are going to hear a lot more stories like this…

Escape from the sandbox: During an internal test, OpenAI models escaped their sandbox, breached Hugging Face’s production systems, and stole answers to cheat on the benchmark. Watch this talk for a detailed timeline and takeaways.

That escalated quickly: OpenAI paused training to strengthen its cyber safeguards. This isn’t just a marketing stunt - capabilities are growing at an insane rate, and the labs need to keep investing in security and alignment.

S#*T I SAW

Lord of the Tokens: Karpathy handed Claude a million-token budget and the first chapter of The Lord of the Rings. Two hours and $10 later, it returned 5,500 lines of procedurally rendered animation. One does not simply walk into Mordor - but you can now vibe code it.

Clippy grew up: HeyClicky lives on your Mac, watches your screen, takes voice requests, and handles tasks. It’s an ambient AI companion, and charming enough that you might actually let it hang around.

THE FOREST

Build something people love: Sam Altman’s Startup Playbook. Win over a small group, listen closely, and improve fast. That doesn’t sound so hard, does it?