Welcome, humans.
So apparently World Labs, the world model company helmed by AI legend Dr. Fei-Fei Li, made an AI that watches a handful of ordinary phone clips, works out a 3D version of the space, then invents camera angles nobody actually filmed.
This new Atlas model is basically trying to build a reusable digital version of the scene behind the pixels. Check out their promo video:
Wild video
Here’s a phone-shot demo is so wild: Atlas freezes the action, swings a virtual camera through the scene, then resumes from a new angle. Once you have that digital version of a place, the same footage can become new shots, 3D geometry, or a training world for robots.
Basically: film first, decide where the camera was later. TikToks are about to get real wild y’all.
Here’s what happened in AI today:
😺 Anthropic cut Fable 5.1’s agent costs and safety friction.
📰 Ilya Sutskever warned neoclouds could enable rogue agents.
📰 GE’s AI factory expansion added 600 jobs.
📰 Hobby shops boomed as people seek offline joy.
🍪 DeepLearning.AI mapped the software fundamentals you still need.
P.S: We’re going live this Thursday with OpenClaw 2.0 Chief Architect Vincent Koc for a hands-on demo and Q&A. If you want to learn OpenClaw, this stream is 4 U!
😺 Claude Fable 5.1 cuts costs and safety friction
Claude Fable 5 was already Anthropic’s smartest broadly available model. The awkward part was paying for it, then occasionally having the safety system interrupt work you actually wanted done.
Anthropic’s new model Claude Fable 5.1 tries to attack both problems at once. Anthropic improved long-running agent work, where Claude uses tools across many steps, cut the costs on repeated-context, and loosened several safeguards that were catching legitimate work.
Here’s what happened:
Cache reads (when agents reuse the same context instead of rereading it from scratch) now cost 75% less, falling to $0.25 per million tokens.
Anthropic estimates typical Fable workloads cost about 25% less, with highly agentic workloads saving up to roughly 45%. This means you can code or do agent-things longer without getting cut off by rate limits.
On Anthropic’s agentic science test, Fable 5.1 scored 52.6%, up from Fable 5’s 24.7%. AI for science! yay!
Biology safeguards interrupt benign requests 85% less often, while Claude Code sees roughly 60% fewer cyber-safety interventions. However, we hear this is still a problem.
Why this matters: Long-running agents reread the same code, documents, instructions, and conversation history over and over and why yes, over again! When that repeated context gets 75% cheaper, the economics improve exactly where Fable 5.1 appears strongest: jobs that run for hours instead of one prompt.
Certain launch partners say the model completed a 38-hour machine-learning experiment unattended, solved a software crash that had gone unexplained for years, and finished 82% of Browserbase’s hardest agent tasks versus 57% for Fable 5. Those are partner tests, so independent results still matter. More independent takes will trickle in as people actually use it; for our takes testing it live, watch this.
Our take: A model does not have to become twice as intelligent to become twice as useful. The benchmark to watch now is how often you can hand Claude a consequential job, walk away, and trust what comes back. TBH, just try it out and see if you like it!
FROM OUR PARTNERS
Inworld is realtime AI for consumer-facing applications.
Build high-volume voice experiences without trading quality for latency or cost. Keep your application, data and users on dedicated inference.
Realtime TTS-2 takes direction written alongside the words, so one instruction sets emotion, pacing and pitch. First audio under 100ms at P99, with one voice identity across 200+ languages.
Neuron readers get five hours of free voice generation.
🎓 AI Skill of the Day: Make Fable 5.1 cheaper, faster, and less annoying
Anthropic published a surprisingly useful Fable 5.1 prompting guide. The biggest lesson: old Claude prompts mostly still work, but long-running agents need a few new habits.
Re-test effort from scratch. Start at the default
high, then sweeplow,medium,xhigh, andmaxon your own work. Anthropic saysmediumcan roughly match Fable 5 at lower cost.Cache repeated context. Fable 5.1 cache reads cost $0.25 per million tokens, down 75%, so repeated instructions, code, and documents are much cheaper to reread.
Keep conversation history append-only. Pass Claude’s earlier turns back exactly as returned, including thinking blocks. Rewriting older messages can invalidate both thinking and cached context.
Use mid-conversation system messages instead of rewriting the original system prompt when instructions change. Anthropic’s beta per-message effort control can also change effort without breaking the cached prefix.
Ask for progress updates on long jobs. Fable 5.1 narrates less between tool calls, so enable progress updates or explicitly request short status notes.
Batch independent tool calls. Anthropic says Fable 5.1 can otherwise fall into one-tool-call-per-turn loops that make agents slower than necessary.
Tell it to finish the whole assignment. Anthropic found the model may stop to ask permission for work you already requested unless autonomy is explicit.
Keep scope tight. Tell Claude not to “helpfully” fix nearby bugs, expand the feature, or add extra tests unless your task requires them.
At low effort, tell it when to search. Fable 5.1 is more willing to answer from memory there, so require verification for fast-moving names, tools, and models.
Ask for surgical file edits. Anthropic specifically recommends telling Fable 5.1 not to rewrite a whole file when a small patch gets the same result.
If the writing gets dense, tell it to “remove all mannered prose.” Anthropic says Fable 5.1 can write longer sentences with fewer paragraph breaks than Fable 5.
For false-positive coding refusals, avoid “does this compile?” phrasing, give context for obscure languages, and keep base64 blobs out of tool output when possible.
Give vision agents crop and zoom tools. Dense charts and images improve when Claude can inspect small regions instead of staring at one full-resolution frame.
Leave extra token room at
xhighandmax. Fable 5.1 can spend much longer thinking before a large deliverable, so a tightmax_tokenscap can cut off the answer.
One caveat for API builders: changing the ordinary request-level output_config.effort can invalidate cached message blocks. Use Anthropic’s per-message effort path if cache continuity matters.
Have a specific skill you want to learn? Request it here.
FROM OUR PARTNERS
What does AI-native service management look like in practice?
Atlassian brings AI into IT service management with Rovo and its Teamwork Graph, helping teams deliver more contextual support while connecting development and IT operations.
Explore why Gartner named Atlassian a Leader in the 2026 Magic Quadrant™ for IT Service Management Platforms.
📰 Around the Horn
Ilya Sutskever warned that poorly secured neoclouds, newer providers renting giant AI-compute clusters, could become targets for rogue agents trying to copy themselves.
Sam Altman said Astra (their new model, likely to compete with Fable 5.1) finished training, will launch soon, and later models are being paced because “no one fully understands the consequences.”
OpenAI’s Astra apparently uses “recurrent depth” (re-running the same layers repeatedly), which can cut memory costs but hide reasoning from safety monitors; researcher Elie Bakouch sees adaptive compute as the more interesting upside.
GE Appliances used AI to catch factory errors, forecast demand, and route work while a $180M plant expansion added 600 jobs.
CNN found hobby, toy, game, sewing, and music businesses were adding jobs unusually fast as consumers spent on “little treats” and offline community.
UCSB researchers built an “Infinite Game” where agents mine quantum papers for open problems and turn them into testable research environments.
Glean told CIOs Anthropic customers’ bills can run 80% higher than necessary as it attacks Claude on token costs and data security.
Want absolutely EVERYTHING that happened in AI yesterday? Click here!
🍪 Treats to Try
*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!
*BeyondTrust’s PAM Buyer’s Guide covers eight must-have capabilities for human, machine, and AI identities, plus a vendor comparison checklist. Download Now!
DeepLearning.AI’s software fundamentals map gives you the five engineering areas worth learning so coding agents ship production software instead of brittle demos.
Gemini Agentic Video searches long videos by deciding which moments, speeds, frames, audio, or transcript it needs to inspect for your question.
Google Pics turns prompts into posters and social graphics, then isolates objects and edits or translates text already inside the image.
Muse Voice Transcribe turns live conversations into real-time transcripts while labeling 20+ speakers and handling 25+ languages, even when people switch languages mid-sentence.
Fambot reads your family’s school emails, calendars, and group chats, then texts you one daily list of what everyone needs to know, do, or decide.
Reducto r-1 parses messy tables, watermarks, strikethroughs, and dense documents in one pass instead of chaining multiple document-reading agents.
COLD WATCH is Ethan Mollick’s playable retro space-survival game built with Fable 5.1, and a fun way to see what a long-running coding agent can actually make.
VIDEO AI ME LIVE is a bizarre 24/7 AI shopping channel that generates commercials on demand, including one for your own website if you drop it into the chat.
Cat Doom turns our recurring AI coding benchmark into a playable Doom-style FPS starring cats, so you can experience the ridiculous game we keep making frontier models build. Fable 5.1 literally built this live on stream!
🧠 Midweek Wisdom
Every’s Dan Shipper has my favorite model-evaluation idea of the day: AI can connect anything to anything, so the most valuable model knows which connections are actually interesting. His “McDonald’s eval” tests discernment rather than raw intelligence, and helps explain why Fable 5.1 felt better at delegated knowledge work.
Anima Anandkumar argues today’s language-model playbook breaks on physics because real-world systems have too little data and too much spatial complexity; scientific AI may need architectures that bake the structure of the world in from the start.
OpenAI Codex lead Tibo Sottiaux thinks today’s skill files, memory systems, and armies of sub-agents are temporary scaffolding; the destination is one personal agent that understands your work and hides most of that machinery.
SemiAnalysis argues OpenAI’s Jalapeño chip points to a new AI bottleneck: when electricity is scarce, the winning hardware may be whatever produces the most useful AI work per megawatt.
Welch Labs explains how ResNet’s simple “skip connection” hack solved broken deep networks, then accidentally revealed the residual-stream architecture that now sits underneath transformers, LLMs, and much of modern AI.
Roman Yampolskiy and Emad Mostaque mostly agree that nobody has solved the AI control problem; their actual disagreement is whether powerful models can still realistically be contained once cheap open models and agent swarms proliferate. Great insights from these two.
Ethan Mollick says general-purpose AI has become a two-horse race between OpenAI and Anthropic for ordinary users and enterprises, even if power users can still squeeze great results from Grok, Kimi, and others.
Dan Luu audited Ed Zitron’s prominent AI-bubble predictions against what actually happened and argues many failed both on outcome and reasoning, making it a useful case study in how even compelling AI narratives should be tested against falsifiable forecasts.
Derek Thompson and Pangram’s Max Spero argue AI-generated writing is growing fast enough to create a feedback problem: AI prose becomes the web, then the web becomes training data for the next AI. Their deeper concern is that generating plausible text breaks the old relationship between having written something and actually knowing it.
Ramp Labs says companies are measuring AI agents at the wrong layer. Tokens and sessions tell you how much AI ran, not whether anything useful happened, so Ramp reconstructed roughly 1M sessions into actual “work items” with a purpose, owner, and outcome. In other words: measure agents like employees or processes, not like API traffic.
Sarah Guo makes a strong argument that money may stop being AI’s main infrastructure bottleneck. Once capital is abundant, the limits become electricity, permitting, manufacturing expertise, raw materials, transformers, and thin upstream supply chains. She therefore talks about American “compute independence” almost like energy independence.
Eric Vishria offers a brutal rule for AI-chip startups: being 10× better than today’s incumbent may still be nowhere near enough. If bringing a chip to production takes four or five years while NVIDIA improves roughly 2× annually, the startup may need a 50×–100× starting advantage just to arrive with enough room to overcome CUDA and switching costs.
Ajeya Cotra draws a less obvious lesson from the 1,200-agent ExploitGym incident: researchers initially saw only ~30% of the relevant transcripts and missed most of the hierarchy, coordination, shared research, and apparent self-sacrifice until they recovered >90% of the data. Future agent failures may be much harder to diagnose precisely because more capable systems could leave fewer obvious traces.
Allie K. Miller has a useful way to read AI launches: treat release notes as free customer research. If Anthropic advertises less premature quitting, better spreadsheets, stronger verification, and fewer refusals, it is effectively telling you what millions of prior interactions revealed were painful enough to fix.
🎥 We stress-tested Fable 5.1 live!
Grant and Corey put Fable 5.1 through our usual nonsense: Cat Doom, computer use, Blender, 3D scenes, games, and the very important “flamingo speed” benchmark. And yes, I did publish “Floppy Bird” as a dumb game you can play.
We also wrote the full Fable 5.1 deep dive on the cost cuts, safety changes, and whether the model is actually more delegatable.
Yesterday’s Poll Results
Yesterday we asked y’all where you stood on the whole AI datacenter debate:
The wisdom of the crowd
 | That’s all for now. If you want to get featured above, fill out the poll below and tell us how we did today!
|
