diwank

this is surprisingly high delta. to make matters worse, reasoning tokens account for the majority of tokens and they are completely opaque so it's hard to tell how much of that is prose or code

i'm not happy with how openai is trying to pit 5.6 sol as a cheaper equivalent to fable here

for one thing, they said that on AA, sol is "within one point of fable" at 58.9 vs 59.9 but don't clarify that the latter is with safeguards where ~8% of the tasks got routed to opus

i'm not rooting for either and genuinely think that the token efficiency and cheaper price are important but this sort of thing just feels disingenuous :-/

where did you find that? weird coz their post announcing this also mentioned Claude Code:

Fable 5 will be available starting tomorrow, Wednesday, July 1, to users globally on the Claude Platform, Claude.ai, Claude Code, and Claude Cowork. For Pro, Max, Team, and select Enterprise plans,1 Fable 5 will be included for up to 50% of weekly usage limits through July 7, after which it will be available via usage credits. We will re-enable access on AWS, Google Cloud, and Microsoft Foundry as quickly as possible.

https://www.anthropic.com/news/redeploying-fable-5

""" It remains unclear whether Anthropic’s engineers are assisting the NSA in active operations. However, one person close to the situation said Mythos would be useful for infiltrating the networks of nations such as China or Iran. """

in order for us to get there, i think we need a standardized api at the os layer for local models so that the os could optimize, batch and safely allocate resources. something like an analog of chrome's local model "prompt" api but provided and managed by the os itself. the user can choose which model they want to primarily use and so on but all of the heavy lifting and continuous batching is done automatically by the os

and device dependent. this is a very tricky thing to get rendered consistently

this is actually a surprisingly rich area of debate in philosophy of mind. see: https://plato.stanford.edu/entries/qualia-inverted/

had a bad experience with pg_search (paradedb) in the past

we have been using pg_textsearch in production for a few weeks now, and it's been fairly stable and super speedy. we used to use paradedb (aka pg_search -- it's quite annoying that the two or so similarly named), but paradedb was extremely unstable, led to serious data corruption a bunch of times. in fact, before switching to pg_textsearch, we just switched over to plain trigram search coz paradedb was tanking our db so often...

also shoutout to tj for being super responsive on github issues!

this is so disingenuous on symbolica's part. these insincere announcements just make it harder for genuine attempts and novel ideas

Angels & Demons anyone?

opus 4.6 gets it right more than half the times

Working on Memory Store: persistent, shared memory for all your AI agents.

https://memory.store

The problem: if you use multiple AI tools (Claude, ChatGPT, Cursor, etc.), none of them know what the others know. You end up maintaining .md files, pasting context between chats, and re-explaining your project every time you start a new conversation. Power users spend more time briefing their agents than doing actual work.

Memory Store is an MCP server that ingests context from your workplace tools (Slack, email, calendar) and makes it available to any MCP-compatible agent. Make a decision in one tool, the others know. Project status changes, every agent is up to date.

We ran 35 in-depth user interviews and surveyed 90 people before writing a line of product code — 95% had already built workarounds for this problem (custom GPTs, claude.md templates, copy-paste workflows). The pain is real and people are already investing effort to solve it badly.

Early users are telling us things like one founder who tracked investor conversations through Memory Store and estimated talking to 4-5x more people because his agents could draft contextual replies without manual briefing. It helped close his round.

Live in beta now. Would love feedback from anyone who's felt this pain! :)

I dont think this is Cerebras. Running on cerebras would change model behavior a bit and it could potentially get a ~10x speedup and it'd be more expensive. So most likely this is them writing new more optimized kernels for Blackwell series maybe?

yeah I agree. this is really unfortunate because it seems that there is something systemic here at play which has become twisted up in a cult of personality and that's made a rigorous scientific investigation very difficult

GitHub Copilot now supports Agent Skills

https://github.blog/changelog/2025-12-18-github-copilot-now-...

Just in time UI is incredibly promising direction. I don't expect (in the near term) that entire apps would do this but many small parts of them would really benefit. For instance, website/app tours could be just generated atop the existing ui.

I am a bit mind boggled by the pricing lately, especially since the cost increased even further. Is this driven by choices in model deployment (unquantized etc) or simply by perceived quality (as in 'hey our model is crazy good and we are going to charge for it)?

Remarkable is really lagging behind on this. I was thinking of finally biting the bullet and writing an app for the Paper Pro. Any ideas/takers?

Google's response:

"Read our statement on today’s decision in the case involving Google Search."

https://blog.google/outreach-initiatives/public-policy/doj-s...

Agreed. The fact that it has any structure at all is fascinating (and super pretty). Could signal at interesting internal structures. I would love to see a version for Qwen-3 and Mistral too!

I wonder if being trained on significant amounts of synthetic data gave it any unique characteristics.

also ettin is a new favorite and a solid alternative: https://huggingface.co/jhu-clsp/ettin-encoder-1b

I'd encourage you to give setfit a try, along with aggressively deduplicating your training set, finding top ~2500 clusters per label, and using setfit to train multilabel classifier on that.

Either way- would love to know what worked for you! :)

It's coming soon! I think this experiment has really taught me a lot about the limits of agentic code assistants, stuff that they're good at, they're insanely good at, and stuff that they're horrible at and cannot seem to overcome. I did write a little bit about how I use Claude Code [1] before I started this project a while back, and I'm planning to finish a sequel pretty soon.

^[1]: https://diwank.space/field-notes-from-shipping-real-code-wit...

yup. I started a fully autonomous, 100% vibe coded side project called steadytext, mostly expecting it to hit a wall, with LLMs eventually struggling to maintain or fix any non-trivial bug in it. turns out I was wrong, not only has claude opus been able to write up a pretty complex 7k LoC project with a python library, a CLI, _and_ a postgres extension. It actively maintains it and is able to fix filed issues and feature requests entirely on its own. It is completely vibe coded, I have never even looked at 90% of the code in that repo. it has full test coverage, passes CI, and we use it in production!

granted- it needs careful planning for CLAUDE.md and all issues and feature requests need a lot of in-depth specifics but it all works. so I am not 100% convinced by this piece. I'd say it's def not easy to get coding agents to be able to manage and write software effectively and specially hard to do so in existing projects but my experience has been across that entire spectrum. I have been sorely disappointed in coding agents and even abandoned a bunch or projects and dozens of pull requests but I have also seen them work.

you can check out that project here: https://github.com/julep-ai/steadytext/

Future robots may learn in their dreams...

So prescient. I definitely think this will be a thing in the near future ~12-18 months time horizon

Same! Guan’s work on sample packing during finetuning has become a staple. His openchat code is also super simple and easy to understand.

I think that’s too harsh a position solely for not being peer reviewed yet. Neither of yhe original mamba1 and mamba2 papers were peer reviewed. That said, strong claims warrant strong proofs, and I’m also trying to reproduce the results locally.

Exactly!

It uses two interdependent recurrent modules: a *high-level module* for abstract, slow planning and a *low-level module* for rapid, detailed computations. This structure enables HRM to achieve significant computational depth while maintaining training stability and efficiency, even with minimal parameters (27 million) and small datasets (~1,000 examples).

HRM outperforms state-of-the-art CoT models on challenging benchmarks like Sudoku-Extreme, Maze-Hard, and the Abstraction and Reasoning Corpus (ARC-AGI), where CoT methods fail entirely. For instance, it solves 96% of Sudoku puzzles and achieves 40.3% accuracy on ARC-AGI-2, surpassing larger models like Claude 3.7 and DeepSeek R1.

Erm what? How? Needs a computer and sitting down.

I'd recommend considering applying to the Thiel Fellowship.

https://thielfellowship.org/faq

Compressed it. Thanks for letting me know :)