$10,000 for filing a legal document in court with 21/23 made up quotes seems ... not much? calmatters.org/economy/tech...

This is one reason why saying "to be successful, you need to work hard" is misleading. The world is full of high talent, strong drive people lacking material macro objectives to get their value to market. They end up in the hands of intermediaries that get a disproportionate cut of the benefits.

Something ugly but true. There are two types of people with strong drives. Macro-level driven people have clear material objectives and care less about the "how" than the "what". Micro-level driven people enjoy finding a space to solve hard problems and learn, for them the "how" is the "what". 👇

Career milestone: Our performance tonight at the Shaky Knees Festival in Atlanta will be my 2,000th live show!

Some TPOT and e/acc folks have started popping in again lately. Hi! Your unhinged posts and chatbot romances are essential to our healthy AI ecosystem, so please check out my old post on constructing your feed to make bsky enjoyable.

- wewoooo - what do you mean the user database got ramsomwared? - weeeeeee - I don't care if you did what you could. I am going to exit the bathroom, return to the meeting and the demo better be working - weeee beep bop beep bop - sigh ...

I am seeing ads of what looks like mobile apps to remotely talk to your coding agent and I am thinking we are so much closer to an R2D2 future than we think.

It would be great if we bring together more folks with opposite ideas. Better if ideas get discussed in a civil way. Better even if the radicals that will try to disrupt this get hellbanned. Even better, they self-hellban into a bsky outragesphere, where they can live happily and not peacefully

I don't have Instagram, so here is a nice day at Central Park. Also the weather was great. Seasons change, not warm, not cold (reminded me of Asturias). The new CP pool was already closed and preparing itself to be an ice skating ring

might be time

Looking for things to be optimistic about the future? Seems like AI will lower the technical barriers empowering those that are more creative (being creative and somewhat technical, even better!).

2035: - when you were small, the program would sometimes crash because there was not enough memory on the computer. - really?? thats lame - it sure was ...

New code review response just dropped:

🤣

Wait, this is a real person?

It took me a long time to get this, but yeah.

IBM Granite-Docling-258M a tiny VLM that fits into the docling toolchain that converts PDF->html/markdown painlessly this addition is now the default for docling huggingface.co/collections/...

One accept and one reject (with a 4445 😔). I do think the criticism was fair, but it's a lot of work to die on the shore ...

We tend to think of context switching as a limiting factor at work, but it is an even bigger problem in "real life". A lot of the stuff we should be caring about we don't do, just because there is a cost of switching from whatever the current task is. (also the attention economy does not help)

First hallucination of Claude Code in the three weeks of work. It wished for a log_stats=True parameter for vLLM's LLM class. It looked so practical! It even had TTFT

If domain adaptation provides a moat against LLM providers is a million (billion) $ question. Watching Cursor's trajectory is probably one of the best ways to answer it

It's great that Claude recognizes :q as a command for exit. It's not great tha It confuse switching vim and tmux splits and kill hours of KV-cached work context

Wut

Definitely one of the biggest wins: Visualization, annotation tools, data re-processing.

TIL, unlike with ssh, if you start mosh on a tmux split (which comes very handy to build a local notebook with remote execution) scrolling will be broken, because of how mosh handles it 😢

IBM's FlowState A Time Series Foundation Model (TSFM), let's say that can be trained on stock price data at one-minute intervals, then used to predict prices at different time scales — such as hourly, daily, or weekly.

When there’s a lot of activity on the site, our dedicated mod team works around the clock to keep you safe and uphold our community standards. If you think there was a mistake, you can appeal by emailing [email protected] and they will get to it as soon as they can.

A/B testing in continuous space.

Crazy random line in the Cursor RL blog post saying they're collecting RL data from real users, updating the checkpoint every 90 to 120 minutes. This would've been unthinkable if you told me it 12 months ago. What a wild year of reasoning models. buff.ly/8LqreS2

One of the things that I found more striking when reading about the Spanish civil war (that's only 90y ago, folks) is how, when it broke out, people were happy to go to the front "to solve the political crisis". Half a million dead later (many in post war executions) the mood was very different.

Coming to terms with the fact that this is my politics now @paperbackparadise.bsky.social

So I guess data centre hardware and software jobs and information security jobs will go up in demand, while demand for refactoring java factories / database schemas, or hacking some custom front end will go down?

My team ran benchmarks using realistic LLM training workloads (llama 70B). CoreWeave infra is the fastest and most stable in the industry. Large distributed jobs survive longer w/o interruptions and recover from faults orders of magnitude faster, reducing wall time, GPU hours, and $$ to train. 1/

oh my god this is just crazy folks github.com/huggingface/...

I got GPT-5 with ChatGPT code interpreter to hunt down the US Census numbers used by the recent Apollo AI adoption rate chart and then recreate that chart from a screenshot and the raw data using matplotlib in Python simonwillison.net/2025/Sep/9/a...

More than 500 researchers from 34 countries published an open letter against the EU's proposed Chat Control regulation csa-scientist-open-letter.org/Sep2025

a good breakdown of the Claude Code agent loop the key: a simple 2-layer architecture “a … single-threaded master loop (codenamed nO) real-time steering (the h2A queue)” (yes, they totally read a de-compiled/obfuscated version, but yeah dedication!) blog.promptlayer.com/claude-code-...

YES HOLY FUCK FINALLY!

I know this is not for everyone, but my gosh, the combination of Project Gutenberg and google docs now having a very high quality AI reader just opens up an infinite world of listening while I do illustrations. Like, books that have been out of print 200 years.

So in the end, I never write the tool. Another one: I just accidentally overwrote a data structure with helpful info, that I can recover from an old file with a different format. Before ... I'd probably give up that data. Now, it took me longer to write this than to ask for recovery code and run it

I was wondering why I am getting slightly less vitriol every time I post here and found out that I am pre-emptively blocked by 146,000 people subscribed to various BlueSky blocklists about AI (and a whole range of other topics). BlueSky is becoming a collection of non-overlapping spaces, I guess.

I never used coding AI-assistants ... directly jumped to assisting AI-coders.

One of the things you notice with coding agents is the compound effect of complexity. Sure, it takes me 10min to write and test a small annotation tool by hand to validate some data. But then it is 20min to change some format or other flag. 30min to refactor the data structures, etc. 👇

Getting to be that time of year.

🤯 this was only 6 months ago

ME, IN TEARS: you can't just say every single part of a computer system is a file UNIX, POINTING AT THE MOUSE: file

Maybe they intend to sneak synthetic employees in it? All owned by OpenAI? Or build the ultimate benchmark "Humanities Last Payroll"

There are exceptional, out of distribution, world sport dominators and then there is whatever freak accident of nature Tony Bou is.

Really capable autonomous weapons systems are right on the horizon. Figuring out your stance on dual use of research isn't really something you can put off anymore.

i.e. pre-training loss is a memorization objective condemned to be always lossy due to model/data size ratio. It's also where most of the magic happens. What did you expect?