Nice work on J-space! Fun coincidence, I explored a very similar technique called logit prisms two years ago: neuralblog.github.io/logit-p… @Jack_W_Lindsey might be worth a look, and a citation would mean a lot if it fits!

People who pick "MoEs are elegant" need to explain themselves lol. MoEs are non-differentiable monsters.

Anthropic should slow down and fix its services.

ok i really like this model architecture, which somewhat proves my intuition that we don't really need a deep encoder for vision/audio modalities, the transformer backbone itself can process the image/audio tokens raw. The vision/audio encoder should be considered to play the same role as the "embedding" layer.

Another day of Google not quite reaching the AI capability frontier.

This isn't actually true, at least with the current version! (That said, Anthropic does offer a context editing feature in the API that lets you clear tool results and/or thinking blocks.) If you want to test it yourself, try this: 1 Ask Claude: "What is 123123123123 x 34534543564654645? No tools, just thinking." 2 Step away for 2 hours to make sure the KV cache gets flushed. 3 Come back and ask: "Give me a 10-word summary for every 200 words in your thinking block before the final answer."

They don't teach you this in school, but Transformers are RNNs. The KV cache is the hidden state.

This looks very impressive, but am I the only one seeing something off here? Do we really need a humanoid robot for this particular task? It feels like someone with a mechanical engineering degree could design a simple machine to do the same thing, right?

In this clip, Dario was talking about *the latest model we release, Claude Opus 4.5* I've already seen a few cases where people post old clips of Dario for various reasons. I'm not commenting on whether Dario is purposely sending *fear-mongering* messages or if he honestly believes in this stuff. I just want to point out that some people are purposely posting old clips of Dario for engagement and/or other agendas.

Someone needs to create a Claude skill to check for hallucinated references and other "meta comments from the LLM" 😅

Cancel your subscription and switch to Codex instead. You'll be doing others a favor too, freeing up compute for the people who actually need it.

I am kind of tired of people's crazy takes on Anthropic. AI twitter is an upside-down world in which the honest people get ridiculed while the liars get celebrated.

Too much of the *hype* around Mythos is focused on cyber. But… Mythos is incredibly good at SWE as well. In fact, cyber is really just a side quest. I'd bet Mythos is great at other tasks too. And I don't think we're prepared for this.

If you want to listen to Claude’s Constitution in your favorite podcast app, I have created an RSS link for it at gist.githubusercontent.com/N… Use your app’s Add podcast by URL / RSS URL option. On iOS Apple Podcasts: Library → Shows → Follow a Show by URL. On Android AntennaPod/Podcast Addict: add/import by RSS URL.

Much better developer experience (Python vs. C++). Much better compile time. Same performance as C++ CUDA.

github.com/NTT123/hill-climb… In Claude Code: /plugin marketplace add NTT123/hill-climbing-skills /plugin install hillclimb@hill-climbing-skills Then: "Run hillclimb on X. Maximize Y from baseline Z."

The dashboard is one self-contained HTML file. State + chart in one file you can email around. Here is a live example, mid-run on ResNet-18 / Flowers-102 fine-tune: ntt123.github.io/hill-climbi…

Six skills that compose: • onboard (interactive setup) • execute (pick top idea, do the work) • verify (run verifier.sh, parse {status, score, notes}) • brainstorm (diagnose + diversify when stuck) • loop (autopilot) • hillclimb (end-to-end orchestrator)

Extends @karpathy's auto-research idea into a generalized Claude Code plugin: optimize any measurable objective with whatever verifier script you write. Hyper-parameter sweeps. Kernel optimization. Accuracy. MSE. TFLOP/s.

I built a Claude Code plugin this weekend: turn any "chase a number" problem into a stateful hill-climbing loop. You give a metric and a verifier script. Claude does the iterating. State lives in a self-contained HTML dashboard. 🧵 below ↓