Guessing we will be soon be asking Claude to draw function stack and logic diagrams to know where to cut the fat.

This so much. We will get scare stories about money wasted and production FUBARs. We will get hype stories about ridiculous development and deploying speeds. The one you end up living depends on you, your peers and how you organize.

Maybe I am doing something wrong, but for me Qwen3/3.5's seem always brittle and inefficient (token-wise) while gpt-oss just works. I have this thing where 3.5 either forgets it uses XML tool format and switches to hermes, in the middle of the trajectory, or breaks into a repetition loop.

Just the story of his name is one of the coolest things of all time

Did not know this!

and this also applies to seeing a "personality" in a model, complicated by the fact that it's a much fuzzier concept. Not saying e.g. Claude and Gemini don't feel like different personalities, they totally do (they feel almost like opposites). I am talking more fine grained stuff.

Instead of improving their predictive ability by getting beter internal states, learn to use the external state for the same purpose

the amount of context that a model needs to solve a task where they always take the right decision is much much smaller than what models use in practice. In fact, thereby lies their 2026 gained "intelligence". Their improved ability to increase their chance of success after each error.

Yes, but large memory also means more attempts at solving a problem, so it's kinda related to intelligence.

It definitely feels like Claude has high variance in its performance which understandably leads to conspiracies ... but my local model does the same. Maybe these things are still high variance, and that's it.

When agentic inspectors start to roam around, it's going to be very funny. "As an ISO 27001 AI inspector, I do not have a mother. Your statement is therefore inapplicable. Your statement is however in violation of section 34(b) of the Interaction Agreement. This incident will be reported."

claude usage quotas are childhood screentime limits coming back to haunt us

The sad state of predictive keyboards (spell check) and on laptop ASR is really puzzling

Astral's original business model was to sell a safe version of the pip python package repo. Is this an agent supply chain security play? I really hope they don't screw uv openai.com/index/openai...

Sloptimism: AI will make bad journalism and fake news so much worse, we will start paying for journalism again.

Minimax m2.7 has just released! I believe this is the first mention from a Chinese lab of early RSI. "M2.7 is our first model which deeply participated in its own evolution.". The benchmark numbers look amazing, around 1 point lower than opus on some benches to a max of around 10 points lower

This is far worse than I imagined. >At a high level, the LLM detection involved watermarking submission PDFs with hidden LLM instructions, which would subtly influence any review produced via an LLM. [this measure was] known publicly [almost the entire review period].

The most disruptive effect of (personal) agents that nobody is talking about is the empowering of individual consumers. The internet entshitified into walled gardens farmed for attention (social networks) and leasing of information (e.g. songs) rather than full ownership, but ... 👇

Flash-KMeans An IO-aware implementation of exact k-means that rethinks the algorithm around modern GPU bottlenecks. Flash-KMeans achieves 30x speedup over cuML and 200x speedup over FAISS. Paper: arxiv.org/abs/2603.09229 Code: github.com/svg-project/...

been feeling kinda stressed lately so I made a little prototype is this anything

And please no try/except and get(variable, <PooA default>)

Also, please don't laugh at me, I have discovered `Ctrl+b ]` pasting later in life.

Some tmux config to avoid ":q"-ing poor Claude set -g pane-border-lines heavy set -g pane-active-border-style "fg=brightred,bg=default" set -g pane-border-style "fg=white,bg=default" set -g window-style 'fg=colour247,bg=colour236' set -g window-active-style 'fg=colour250,bg=black'

I needed the hint

TIL the solution to the tmux inception problem i.e. you opened a tmux inside of a tmux, is just pressing Crlt+b twice to act on the inner tmux

Interesting read despite this disclaimer! > if you are doing language models, honestly, none of the rest here applies to you. You are probably going to use GRPO anyways, so my work here is done. Have fun, and don’t build ASI before I get a tenured faculty position please.

Smart ways of distributing cognitive off-loading in time will result in far cheaper average tokens

Addendum: Third should always be "exceeded expectations", but can include (and it's probably the most frequent case) that, while exercising agency, we go through uncertain moments where we fear defeat and thus we lower our expectations, so that achieving the final goal actually exceeds them.

I mean just from a psychological point of view. No idea about what economical system would survive the death of white collar jobs, but anything I can imagine looks pretty bad.

"playing only" seems more realistic than UBI as something that would compensate for the death of white collar jobs.

I mean that they don't like teaching/mentoring. You can definitely do this in a corpo job too! But if you like so, you can take an IC route. Note that the demand for interpersonal skills will still be higher in corpo world than academia.

So there was no knot. Thankfully we have people with the skills to straighten this rope

i think it's a great paper, it's like a blessing to me it's making so many ideas sharp that were fuzzy before but yeah he's basically saying in 2.2 that phenomenology is instantiated by the special causal powers of our neurology

Humans have certain prior beliefs about how hard some things may be to explain/understand. I am not free from that. If someone is explaining to me how a quantum computer works (which is very cool!), I expect it to be hard. 👇

This whole thread is pretty amazing. Some profs. seem to want corporate jobs and they don't know it x.com/i/status/203...

So what's AI for growth? Well, AI for advancing science and engineering, forecasting and planning, system regulation, etc. Will that solve the likely upcoming power concentration problem? Nope. But it will make us less desperate. Collective intelligence drops like a brick if we are too scared.

Some AI automation scenarios are starting to look very scary. What can we do? (Re)inventing some SaaS business, even if it is the most inspired idea, will just transfer wealth. If no one has money to spend, it will fall with everyone else. We need growth, and we should use AI for this.

🧵New paper: "Lost in Backpropagation: The LM Head is a Gradient Bottleneck" The output layer of LLMs destroys 95-99% of your training signal during backpropagation, and this significantly slows down pretraining 👇

I was not finding the words, but this sums it perfectly.

is a calculator a type of cognitive off-loading or no? i am trying to understand the objection

I see this weird pattern where every post I make that has a typo gets x10 the average attention

As well know, this will be goodharted and stop being an issue in one two model generations

If you ever suspect you’re talking to an #AI deepfake avatar bot, ask them to count from one to five on their fingers. (Veo 3.1, Sora 2, Kling 3.0)

(and N=3)

Android enabling (beta) AppFunctions a MCP equivalent to allow Apps to be controlled by agents android-developers.googleblog.com/2026/02/the-...

Ok this is more serious auto-programming (if it works) github.com/RightNow-AI/...

There is this little known secret that tuning your seed N times increases your chances of success from p to 1 - (1-p)^N a weakling model with 50% hit rate becomes a 87.5% beast with this amazing trick

This weird sensation when you debug something by talking to someone, iterating over possible reasonable causes until you find it, and that someone is a machine.