Notes By Lex (@notesbylex_) profile picture
Joined August 2024
Photo by Notes By Lex (@notesbylex_)Photo by Notes By Lex (@notesbylex_)Photo by Notes By Lex (@notesbylex_)Photo by Notes By Lex (@notesbylex_)Photo by Notes By Lex (@notesbylex_)
Claude Fable 5.2: some small steps, some giant leaps. About 20% cheaper per token than Fable 5.1, but token-hungry: a crazy $26.54 per task on Artificial Analysis. On my MNIST weight-writing test it beat Opus 5.5, 81.4% to 71.0%. notesbylex.com/claude-fable-…
5
Fable 5.2 one-shotted my nephew's dream game from a bunch of drawings, and it's a lot of fun: a 12-level, 18-character, four-player beat 'em up. notesbylex.com/fable-5-2-one…
8
Frontier LLMs can read, write and edit neural network weights. In my blog post, I explore the claims of Weight-Space Programming and how the results work. I also tested game identification with eight other models. notesbylex.com/weight-space-…
12
Continual learning approach that uses memory consolidation + RL “dreaming” approach: generate synthetic examples and fine-tune on the highest-signal ones. Cool idea. research.google/blog/the-sle…
Notes By Lex retweeted
There is something deeply wrong with a political system that allows one trillionaire to buy his own personal U.S. Senator.
There is something deeply wrong with Talarico
10,023
26,193
993
223,288
5,876,347
Wrote some notes on a new paper from Meta and friends called Context Language Models. They basically hand over context management to the model: Context As A File; with this newfound freedom, the model developed some novel context-management strategies. notesbylex.com/context-langu…
7
Wrote some notes on Google's long-awaited new frontier model: Gemini 4 Argon. It can write up to 1M tokens in a single response (well...in batches), and launches at GPT-6 Sol prices. Ran a few other audio, video and coding tests in my post: notesbylex.com/gemini-4-argo…
4
Wrote some notes on Meta's new Muse Wonder frontier model. It's close to Claude Opus 5.5 on many benchmarks at the same per-token price, but uses fewer tokens in my tests. It also takes audio input and almost perfectly transcribed a two-person chat. notesbylex.com/muse-wonder
3
I wrote a post about building an agent harness from scratch and testing it on a challenge Ian Goodfellow said in a 2019 interview would convince him we've achieved "real AI". It's amazing how fast things have progressed. notesbylex.com/building-an-a…
1
9
Great article on anti-patterns in software blogging. Definitely been guilty of many of these at times. refactoringenglish.com/blog/…
4
Shared a few notes on Anthropic's new Claude Haiku 5.5 model. notesbylex.com/claude-haiku-… TL;DR - priced the same as GPT-6 Luna per token, although it uses about 1.5-1.7x as many tokens (same tokeniser as Opus 4.7). Token costs are also 5x higher when the prompt is over 100K tokens.
3
Opus 5.5 - posted a few notes and a demo on my blog. Major difference from previous Claude versions is that thinking cannot be disabled, as usual substantial improvement on benchmarks. notesbylex.com/claude-opus-5…
8
Jev is a cool project, but it's kinda surprising how enthusiastic people are about classification models again.
3
After about a year of wearing a smart ring, I've decided to throw in the towel. A short essay about the downsides to smart rings: notesbylex.com/giving-up-on-…
5
I wrote some notes and a code examples about Stateless MCP servers notesbylex.com/what-is-state…
4
LLMs are still the wrong tool for tabular data "no amount of prompt polishing aimed at for- mat, precision, or batching should be expected to fix high-dimensional in-context prediction, and purpose-built tabular models remain the right tool" arxiv.org/abs/2608.02412
4
Good article. This approach is something I've been thinking about recently too. The trouble is that our brains have an innate desire to conserve energy, and AI agents (and LLMs in general) make it very easy to give in to that impulse. ankursethi.com/blog/prevent-…
3