Pablo Felgueres

@rdotapps Using this water flow sensor. With temperature delta and flow rate you can then calculate cooling power

@pfelgueres How are you measuring flow rate in the loop, or is the dash mostly using temperature delta and input power?

Every watt that goes into generating tokens turns into heat I made this rig to learn how liquid cooling works. The heat source is an aluminum block with a resistance inside and the loop is made out of PC parts, arduino, and sensors. Also built a dash that shows the energy balance in real time. What I found interesting is that the AI buildout is targeting 25 GW / yr for chips which means that there’s an equivalent industrial buildout happening around thermals: quick disconnects, cold plates, pumps, radiators, etc Incredibly positive for manufacturing and the trades

Today we’re launching our Legal Agent Benchmark, an open benchmark for long-horizon legal agents. This is a culmination of a tremendous amount of effort at Harvey that followed the coding agent inflection point earlier this year. 24 practice areas —> 1250 tasks —> 75,000 rubric criteria The two biggest reactions we’ve had from research partners are: 1. I had no idea law was this complex 2. There’s so much work to do — from evals & benchmarking, to harness optimization & agent building, to post-training & more. Excited to be partnering with a stellar crew of frontier labs, agent builders, model trainers, researchers, and most importantly the open-source community to maximize agent quality for the legal industry! And the best part, LAB is fully open-source: github.com/harveyai/harve…

@marlouiise Neat trick!

A fun hack in Claude is to ask it to reveal the underlying grid structure so you can see where things are off

@SIGKITTEN Ergonomics on point

trying out this card keyboard, kinda cool

@rtwlz Avoid street cleaning parking tickets streetcleaningparking.com

I reverse engineered the San Francisco parking ticket system. I can see every ticket seconds after it's written So I made a website. Find My Friends? AVOID THE PARKING COPS.

@karpathy very similar to Minky's FRAMES idea (1974)

We're missing (at least one) major paradigm for LLM learning. Not sure what to call it, possibly it has a name - system prompt learning? Pretraining is for knowledge. Finetuning (SL/RL) is for habitual behavior. Both of these involve a change in parameters but a lot of human learning feels more like a change in system prompt. You encounter a problem, figure something out, then "remember" something in fairly explicit terms for the next time. E.g. "It seems when I encounter this and that kind of a problem, I should try this and that kind of an approach/solution". It feels more like taking notes for yourself, i.e. something like the "Memory" feature but not to store per-user random facts, but general/global problem solving knowledge and strategies. LLMs are quite literally like the guy in Memento, except we haven't given them their scratchpad yet. Note that this paradigm is also significantly more powerful and data efficient because a knowledge-guided "review" stage is a significantly higher dimensional feedback channel than a reward scaler. I was prompted to jot down this shower of thoughts after reading through Claude's system prompt, which currently seems to be around 17,000 words, specifying not just basic behavior style/preferences (e.g. refuse various requests related to song lyrics) but also a large amount of general problem solving strategies, e.g.: "If Claude is asked to count words, letters, and characters, it thinks step by step before answering the person. It explicitly counts the words, letters, or characters by assigning a number to each. It only answers the person once it has performed this explicit counting step." This is to help Claude solve 'r' in strawberry etc. Imo this is not the kind of problem solving knowledge that should be baked into weights via Reinforcement Learning, or least not immediately/exclusively. And it certainly shouldn't come from human engineers writing system prompts by hand. It should come from System Prompt learning, which resembles RL in the setup, with the exception of the learning algorithm (edits vs gradient descent). A large section of the LLM system prompt could be written via system prompt learning, it would look a bit like the LLM writing a book for itself on how to solve problems. If this works it would be a new/powerful learning paradigm. With a lot of details left to figure out (how do the edits work? can/should you learn the edit system? how do you gradually move knowledge from the explicit system text to habitual weights, as humans seem to do? etc.).

live here: reads.pablofelgueres.com/bookshelf

Made a tool that extracts Kindle notes to use with LLMs. So far: • Gets highlights, notes, and metadata • Keyword search across books • LLM explains highlighted sections I want an effortless way to revisit reads, search, and find evidence in favor or against ideas. There's so much friction on Kindle that this space seems very unexplored.

The information density per token in game development is really high. A solo dev can now one shot years of learnings from specialized fields, overnight getting polymathic skills in graphics, physics, psychology, etc. It’s all in the weights. I’d bet the one-person 1B company will be a game.

@dwarkesh_sp Permutation City en.m.wikipedia.org/wiki/Permutati…

Favorite novels?

@zekedup 80s tech, many closing down because can’t compete with PV. Here’s an amazing scene from gattaca

@StudioMilitary @ErikVoorhees lol love your nous work! fwiw the mask is super cool but the shape on top makes it look like a computer chip stock icon

@pfelgueres @ErikVoorhees I did the Venice logo lol

Venice.ai is hiring two roles immediately to help build the world's leading private & uncensored AI platform... + Sr. Backend Engineer + DevRel Lead for Venice API 1/

A starter pack to the US — its spirit, structure, and sources of wealth

released! apps.apple.com/us/app/street-… avoid parking tickets - locate from gps or search by address - share parked location and reminders - get notifications to move - updated with 2024 schedules

@KTVU Dmed

@pfelgueres Hi Pablo, We'd love to have you on as a guest to talk about the "Street Cleaning Parking" app that you and your girlfriend, MaryLouise Howell, created.

wrote this app last weekend and blew up on reddit. if you get parking tickets check it out it's useful using @sfgov data

@rsms @sfgov Means a lot! Been a long time fan of your work :-)

@pfelgueres @sfgov Great idea! Will definitely save me money in the future :•)