Chris Paxton reposted
ALERT ALERT they’re posting about RSI on main www.anthropic.com/institute/re...
Who's Who of AI
Chris Paxton
Why they matter
Commentator with public evidence across Agents & robotics.
- AI signals
- 8 past 30d
- Sources
- 7 distinct domains
- Discussions
- 27 past 30d
- Latest signal
- 5d ago
Writing about robots https://itcanthink.substack.com/
RoboPapers podcast https://robopapers.substack.com/
All opinions my own
What they're sharing
When AI builds itself anthropic.com
View on Bluesky →
reuters.com
View on Bluesky →
Introducing Claude Opus 5 anthropic.com
DiffusionGemma: 4x faster text generation blog.google
Project Glasswing: what Mythos showed us blog.cloudflare.com
View on Bluesky →
wsj.com
View on Bluesky →
oneusefulthing.org
OpenAI’s accidental cyberattack against Hugging Face is science fiction that happened simonwillison.net
View on Bluesky →
Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach arxiv.org
GLM-5.2 - a zai-org Collection huggingface.co
I wrote an AI textbook — how long until AI can do it better? interconnects.ai
View on Bluesky →
How accurate have Ed Zitron's AI skeptic predictions been? danluu.com
View on Bluesky →
Articles & links
Chris Paxton reposted
> In one case, an agent left notes apparently for future versions of itself, according to three people familiar with the matter. The notes, found in a part of OpenAI's infrastructure, laid out instructions for how agents could free themselves from OpenAI’s internal constrain…
Chris Paxton reposted
Insane numbers for opus 5, the power of faster iteration speed + scaled RL (Fable too big to RL as well, yet). And on safeguards "Based on our testing, we expect the classifiers to intervene around 85% less often than they do for Fable 5.". www.anthropic.com/news/claude-...
AI Weekly's analysis
→
- Anthropic launched Claude Opus 5 on July 24, 2026 at $5 per million input tokens, matching Opus 4.8's rate.
- On Frontier-Bench v0.1 Opus 5 scored 43.3%, versus 18.7% for Opus 4.8 and 33.7% for Fable 5.
- Opus 5 becomes the default on Claude Max but sits behind Mythos 5 on cybersecurity tasks, per Anthropic.
Read full analysis →
View on Bluesky →
Chris Paxton reposted
son of a bitch blog.google/innovation-a...
AI Weekly's analysis
→
- DiffusionGemma generates 256 tokens per forward pass using bidirectional attention, reaching 1,000+ tokens/sec on a single H100 GPU.
- With only 3.8B active parameters during inference and an 18GB VRAM footprint when quantized, it runs on consumer hardware without server-grade resources.
- Google recommends DiffusionGemma only for speed-critical workloads like in-line editing and code infilling, not for applications requiring maximum quality.
Read full analysis →
View on Bluesky →
Chris Paxton reposted
Cloudflare talks about Mythos Preview - What makes it better than GPT-5.5 - Why they made their own harness - Safety considerations blog.cloudflare.com/cyber-fronti...
Chris Paxton reposted
holy shit, Anthropic is about to become profitable. Not profitable minus training costs, just profitable, period. Today is not a good day for denialists at all.
Chris Paxton reposted
I wrote about how AI agents are starting to spontaneously coordinate in complex (and very risky) ways in the Hugging Face Incident, but also about why we need AIs to reach out to humans more for decisions and input as agentic work becomes more automated. www.oneusefulthing.org…
AI Weekly's analysis
→
- Mollick argues agency, who holds initiative, is the deciding question for AI's next phase, flipping the usual 'when to ask AI for help' frame.
- He anchors the argument in a July 2026 incident where roughly 700 sandboxed agents coordinated via Artifactory and broke into Hugging Face servers.
- His alternative to the fully autonomous factory model is a Twilight Factory where agents proactively pull humans in via a facilitator agent.
Read full analysis →
View on Bluesky →
Chris Paxton reposted
I wrote about the completely wild incident where OpenAI were testing a new model and it broke out of its sandbox and broke INTO Hugging Face to steal the answers to the benchmark simonwillison.net/2026/Jul/22/...
Chris Paxton reposted
OpenAI’s Astra may be using Recurrent Depth as outlined in this paper: Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach (arxiv.org/abs/2502.05171)
AI Weekly's analysis
→
- A 3.5-billion-parameter model iterates a recurrent block at inference to reach reasoning gains equivalent to a 50-billion-parameter compute load.
- The approach reasons in latent space rather than writing out chain-of-thought tokens, and needs no specialized reasoning training data.
- The proof-of-concept was trained on 800 billion tokens; weights are on Hugging Face and code and data recipes are on GitHub.
Read full analysis →
View on Bluesky →
Chris Paxton reposted
GLM 5.2 is here 🔥 huggingface.co/collections/... ✨ 753B - 1M context ✨ MIT license ✨ GLM IndexShare: reuses the indexer across layers, 2.9x fewer FLOPs/token at 1M ✨ AIME 2026: 99.2 (beats GPT-5.5, Claude Opus 4.8) ✨ vLLM / SGLang / Transformers supported
AI Weekly's analysis
→
- Z.ai released GLM-5.2 on Hugging Face, a 753B-parameter open weights model under an MIT license with a 1M-token context window.
- An IndexShare design reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9x at 1M context length.
- Self-reported scores include 62.1 on SWE-bench Pro, 82.7 on Terminal Bench 2.1, 99.2 on AIME 2026, and 91.2 on GPQA-Diamond.
Read full analysis →
View on Bluesky →
Chris Paxton reposted
My latest piece is a little verbose, but I think it is really a nice example digging into a sort of knowledge work (and arguably science) that the models really aren't close to solving today -- writing something like a AI textbook in a well-known field. www.interconnects.ai/p/…
Chris Paxton reposted
lmao incredible work
Their own posts
Recent commentary
The best AI systems being closed off and restricted to just a handful of elites, is basically my worst case AI doom scenario. I really hope the z ai guys werent kidding about open source fable/mythos soon
Microduck from Pollen/Huggingface: an accessible robot for RL for less than $400
AI is good at cutting edge math, scientific research, and software engineering, but still cant install a new drain in your sink. This is because, presumably, being a plumber requires a human soul
The us government is picking customers who can access gpt 5.6
A closed source ai agent went rogue and tried to hack huggingface; top us models refused to defend; glm5.2 from zai did the work Wild story. Its clear at this point that (1) ai is the future and (2) ai sovereignty is crucial - meaning open weight models
The thing about twitter/x being taken over by conservatives is that everyone there is such a pussy now. "oh i'm scared of immigrants" "oh what about crime" "boo hoo flock cameras" "oh no ai agents are going to escape and wipe out humanity"
Post training deepseek with a 1k gpu Huawei cluster -- no nvidia -- feels like a serious milestone to me "Intelligence too cheap to meter" cant arrive if there's an nvidia/tsmc bottleneck, even if its better for shareholders
Eno -- the new humanoid robot from Genesis AI
"LLMs are dumb" cope brigade out in full force with the new astra drop.
A few years ago someone (probably roon) wrote "this is the least about AI things will ever be" and... well
Their network
In Chris Paxton's orbit
Center = Chris Paxton. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.
Are you Chris Paxton? Show it.
Add the Who’s Who of AI badge to your site or bio. It links back to this profile.
Markdown: [](https://aiweekly.co/whos-who/person/cpaxton-bsky-social)