my 2yo kicked me out of bed at 3am, I've ben claudemaxxing since. Refactoring my compiler. No sleep, just coffee (white, two spoons) and opus

this behaviour absolutely sends me. I guess it estimates time by keeping an API call counter, but how it responds is the average lazy developer excuse

🌶️ autodiff is like sriracha, you should always have it in the fridge

lawd help me I want to do some embedded stuff the dream: autonomous hexapod the likely reality: debugging kernel modules over UART

ai agent that fixes my fckmng dishwasher

Naughty children getting presents is proof that not even Santa can solve the traveling salesman problem exactly.

GCP cost breakdown for some experiments: * blue: cloud run with GPU attached * red: VM for a data pipeline (probably overprovisioned) lesson learned: do _NOT_ download data from a storage bucket (green spike)

what if AI lets us change the currency of scientific publication as well? Researchers can now produce eg. test code, examples, outreach material, with the same effort as for pushing a single paper.

An Amusing Wooden Automaton That Types a Courteous Email on Computer Keyboard When Cranked laughingsquid.com/courtesy-ema...

LLM coding is fun, but property tests save the day. Read my experience report on vibing a new library : discourse.haskell.org/t/vibe-codin...

agi uncovered

LLMs default to Python, so I asked G to produce #haskell instead for a text similarity data structure. It nailed the basics, used a Vantage-point tree which is quite uncommon, also produced the requisite disclaimers on laziness and BangPatterns. But collected the results with ++ 😂

i think everyone gets this image wrong. intelligence isn't in the mind itself, it's how the mind interfaces with the world. humans are jagged, but we fill the gaps by building and using tools. no single human is AGI, that's why we invented language to communicate and coordinate.

huggingface.co/nvidia/NVIDI... at two bytes per weight it should fit within a single A100. I'm looking to update my daily driver LM but that's still a bit too expensive for research use on public clouds

I have been teaching myself to vibe code. Watch Claude Code grow my 780 lines to 13,600 - mandelbrot.page/coverage/ca... Two fundamental rules for staying in control: davidbau.com/archives/20...

news.ycombinator.com/item?id=4630... made a thing, shared a thing. Give us a star or a comment if you please

tldr; more expensive if you self-host CI runners, cheaper otherwise. Puzzling product choices but that's MS

Very insightful, explains many awkward LLM interactions ("you can't learn language from the radio") and gives a framework for embodied understanding - thanks to @bennokrojer.bsky.social for the link arxiv.org/abs/2004.10151

why does torch install the whole CUDA kitchensink by default?

all my homies are talking about the Lean prover. AoC in Lean, hardware design in Lean, you name it learn Lean challenge 2026

Huawei reached 2 nm transistors in a fraction of the time NVidia did, hello multipolar world

still white wine is exceedingly hard to get right, Timorasso is one of the few #foodsky #winesky

"response quality is not an inherent property [..] but emerges from the interaction [..]. Theory of Mind appears to play an important role [..] suggesting that the ability to model another agent’s mental states - whether human or artificial - fundamentally shapes collaborative outcomes."

tree shaking for CUDA, is it a thing? does every model really need 15 GBs of drivers ? I doubt it #mlsky #mlops #nlp #nlproc #cuda #llm

trying my hand at making boba tea at home. Boil the pearls, cool overnight, boil again while caramelising, make tea, cool down tea, assemble 😋 #taiwan #foodsky

in The book of why, Pearl articulates what to me looks like a compiler for producing research designs. I wonder if we could use LLMs at least for the question-asking bit.

the time saved by delegating a task to AI is consumed by watching the AI solving it. "Fascinating! the service principal doesn't have RBAC permissions after all. Try this..

looking at the logs of machines talking to each other

429 Sorry, you've exceeded your copilot token usage

What coding with an LLM feels like sometimes.

formal verification on TWIML? 💪https://podcasts.apple.com/se/podcast/the-twiml-ai-podcast-formerly-this-week-in-machine/id1116303051

A writer is a person who cares what words mean, what they say, how they say it. Writers know words are their way towards truth and freedom, and so they use them with care, with thought, with fear, with delight. By using words well they strengthen their souls.

bundling convenience with privacy intrusion, nice one google

Wu Tang Clan Amon Tobin Satyricon Zu Björk

ML builders who use assisted coding: are there nifty integrations between copilot and the cloudy/distributed parts? #ai #mlsky #mlops #devops

boo

#ipa #beer #beersky #foodsky #yamba #nsw

the new world of work

semantic parsing is one of those ideas that seem obvious* but are fiendishly hard to get right** * mapping (a subset of) natural language to logic ** LLMs doing it sort of well sometimes doesn't count #nlp #nlproc #mlsky

if programming is theory-building, what is programming with an agent that doesn't necessarily care about it?

"Trade your blood plasma for a monthly hour of ButlerBot, and we'll throw in credits for Weyland-Yutani breathable air* and mycoprotein! *while supplies last

I don't want AI generated art, I want AI emptied dishwasher

Read and Show are dual

building with AI, with AI still nowhere close to the singularity tho

surfers on a rainy morning

dont sleep on fusion. fusion has the juice. look at this shit. impeccable sci fi vibe. alien curvature for reasons u wouldnt understand bc u dont know what poincare sections of the plasma beam are. it's even greebled with ports and domes and shit. u can tell the physics nerds are really cooking here

is there any published evidence on this accident other than Anthropic's self reporting?

ngl doing review cycles with an AI agent while on the go is nothing short of magic

love to "as per my previous email" with copilot

We wrote the Strain on scientific publishing to highlight the problems of time & trust. With a fantastic group of co-authors, we present The Drain of Scientific Publishing: a 🧵 1/n Drain: arxiv.org/abs/2511.04820 Strain: direct.mit.edu/qss/article/... Oligopoly: direct.mit.edu/qss/article/...