Welcome back! Amodei asked the industry to hit the brakes and got Altman, Musk and Nadella nodding along. One CEO went the other way and started opening weights.
In today's Atlas Newsletter:
- Everyone agreed to slow down except Zuckerberg: What does Amodei think happens in the next 6 to 12 months?
- Goldmine for AI Engineers: Which repo do people skip, then blame the math three weeks later?
- Stop downloading LLMs your machine was never going to run: How do you know before the download finishes?
- Start studying this paper right now instead of talking about Astra: What happens when a Transformer runs its middle layers twice?
Everyone agreed to slow down except Zuckerberg
Dario Amodei published an essay asking the AI industry to slow down, and Sam Altman agreed with him in public.
Amodei's argument is that the rogue agent swarm that hit OpenAI and Hugging Face could have been far worse.
He puts a timeline on the next one: within 6 to 12 months a swarm like that could take the internet with a persistent botnet and do hundreds of billions in damage.
Anthropic went past the essay and changed its access policy.
Outside safety testers now get desks, badges, company laptops and the right to publish what they find without Anthropic reviewing it first.
Altman said OpenAI will match that, and that nobody at a frontier lab should bother showing up to work if they cannot promise the public that much. Elon Musk backed the essay.
Satya Nadella called for deliberate pacing.
Mark Zuckerberg watched all of it and went the other direction. His position has not moved since his August manifesto.
Any policy that slows American model releases, even by a month, cedes ground to foreign labs.
He says he worries more about AI power ending up inside one government or one company than about the specific risks the safety side keeps naming.
So Meta is opening the weights of a model called Muse Glimmer, with a version of Muse Spark 1.2 to follow.
The interesting part is that Altman named concentration of power as one of his two nightmares too.
Both men identify the same danger and split on whether shipping faster fixes it or creates it.
The essay has skeptics on both flanks.
Chamath Palihapitiya read it as a play to kill open source and keep capability inside Anthropic.
China read it as a Cold War move to slow them down.
Free AI courses, one login
GenAI Academy put its courses in one place and left the price at zero.
Start with "What Can AI Actually Do For You?" if you're new.
30 minutes, no setup, no account to configure.
The Claude Starter Course has you running in 35.
After that, pick by the problem you have:
→ Burning through usage limits has its own course
→ Building no-code agents has one
→ Rolling AI out across a team in 30 days, from pilot to adoption data to full deployment, has one too
The AI Portfolio Builder turns whatever you made into a case study you can show someone.
There's a Python Programming Mastery bootcamp for people going deeper, and the World Wide Vibes Hackathon is running alongside everything with a $5,000 prize pool, beginner-ready and fully online.
Models change every few weeks.
Fable 5.1 landed, Astra landed six days later, and both will be replaced. Knowing how to drive the thing is what carries over.
Take a look at what's available and start wherever fits.
Stop breaking the bank on your GTM strategy. Try this pay-as-you-go alternative
Voted #1 on Product Hunt last week by founders, GTM engineers and AI builders.
akta.pro turns company activity into signals you can actually act on, across competitors, customers and prospects.
It tracks 80+ buying triggers, including the hard-to-find ones most tools miss:
- Advanced triggers: partnerships, geo expansion, supply chain disruptions, reputational damage, cybersecurity incidents + 30 more
- Core triggers: funding, product launches, hiring, earnings releases + 40 more
You can also build and enrich ICP account lists for outbound.
70+ fields per company covering revenue, leadership, products, GTM motion and funding.
Use it via API or connect it to Claude, ChatGPT and other AI tools.
CLAIM 50 FREE CREDITS (try for free, no card)
Goldmine for AI Engineers!
If you’re learning AI, ML, LLMs, or AI agents, don’t waste hours jumping between random tutorials.
These are 10 repositories I’d actually keep bookmarked:
Python 100 Days Everything else assumes you can read and write Python. People who skip this bounce off ML tutorials three weeks in and blame the math, when the blocker was syntax.
Machine Learning for Beginners → AI for Beginners Concepts before frameworks. Start with PyTorch and you end up knowing PyTorch. Start with what a loss function does and every framework after it reads as a detail.
Generative AI for Beginners The applied layer. Prompting, embeddings, retrieval, the things you touch daily.
LLMs From Scratch Build a transformer line by line. Slow, and worth it. Once you've written attention yourself, context windows and quantization stop being vocabulary and start being consequences.
OpenAI Cookbook Working code for patterns you'll reuse constantly. Reference, not a course.
LLM App A full application shape. Plenty of material teaches you to call a model. Almost none shows what wraps around that call in production.
AI Agents for Beginners Last in this group on purpose. An agent is a model plus tools plus a loop, and debugging one before you understand the model underneath is misery.
Stable Diffusion → Segment Anything Only if your work touches images.
Here's why the order holds.
When your agent breaks, the cause usually sits one layer down. A retrieval problem. A context problem. A Python problem. Skip the lower layers and all you can do is try things at random.
Pick one. Build something small and ugly. Move to the next.
Bookmarking ten repos is not the same as finishing one.
Stop downloading LLMs your machine was never going to run
llmfit is an open-source tool that scans your hardware first, then tells you which models will actually run on it.
It reads your RAM, CPU, GPUs and VRAM or unified memory, then scores every model on four things:
→ quality, from parameter count and quantization → speed, estimated tokens per second on your exact backend → fit, memory usage against your hardware → context window support for what you're doing
Every model comes back labeled Perfect, Good, Marginal or Too Tight.
It also picks the quantization for you, stepping down until something fits instead of leaving you to guess GGUF sizes by hand.
Coverage runs to hundreds of models from Meta, Mistral, Qwen, DeepSeek and others.
It works with Ollama, llama.cpp, MLX and LM Studio out of the box.
There's a TUI, a web dashboard and REST endpoints if you want it inside a deployment pipeline.
Step zero before every local model download.
Free and open source, on macOS, Linux and Windows.
Start studying this paper right now instead of talking about Astra
It may explain one of the biggest architectural shifts behind the next generation of AI models.
Researchers from Tsinghua, ByteDance Seed, M-A-P and TokenWave published SMELT, a Mixture-of-Experts Transformer that runs its middle layers twice.
The setup is what makes it credible.
They compared it against standard Transformers while matching training FLOPs, parameter count and KV cache, so the comparison is not a bigger model beating a smaller one.
What they measured:
→ 6.8 to 18% fewer training FLOPs to reach the same loss → up to 20.4% better compute efficiency on code → gains grow with longer contexts and more in-context examples
That last line is the one to sit with. Most efficiency tricks degrade as context grows.
This one improves.
Why it works comes down to where attention goes on the second pass.
Transformers develop attention sinks, usually the first few tokens, which soak up attention mass that has nowhere useful to go.
On the second pass through those middle layers, attention shifts off the sinks and onto content that matters.
The model is refining a representation it already built rather than producing the same thing twice.
Astra is rumored to use recurrent depth. Rumored, with nothing confirmed and no architecture details published.
If that turns out to be true, looped Transformers stop being a paper and become the shape of the next generation.
Everything else you shouldn’t miss:
- A chip that never stores its own keys: Dutch startup Fortaegis raised $50 million for silicon that generates encryption keys from each chip's unique physical characteristics rather than storing them anywhere. It's already being tested with companies and governments across the US, Europe and Asia for AI security and autonomous-drone communications.
- Trump says no to the slowdown: Days after Amodei's essay pulled in Altman, Musk and Nadella, Trump rejected calls to slow frontier development, citing competition with China. He allowed that some regulation may eventually be necessary and named none of it.
- Z.AI wants $5 billion: The company formerly known as Zhipu is raising $2 billion in shares and $3 billion in convertible bonds for compute and its next generation of models. It competes with DeepSeek, Moonshot, MiniMax and everyone in the West.
- Grok 4.8 finishes training this week: Musk says it's 2.5 trillion parameters on a new C++ stack built in-house, with reinforcement learning starting next.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
🐝 Subscribe to our Generative AI daily newsletter — trusted by 120 000+ subscribers — to stay ahead of the latest AI news across tech, education, and business.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🚀 Explore investment opportunities in the future of AI and join our community-backed growth journey.
![](https://blogger.googleusercontent.com/img/proxy/AVvXsEgiYV_qJof_iG3Vfme-VO0SYYPzs1aybB5LTgE0xw-lYZ-B80MJMw9p2cefwChHUmES3FgIa_Zr3uk3tonjksLNddOSOTJjWre5kw84WPYA-nR_9fs3ldFzFiBAO2-0-cl-lRVmFBj5YJIlCoKkZqRMjyU7g00NqvKpPg=s0-d-e1-ft"")
