vitomd

Software Engineer. I like #rails and #javascript. My site:

[Andrej Karpathy @karpathy]

[New Talk] Pleasure to come by Microsoft BUILD this year and give a talk on "State of GPT". Goes through the GPT Assistant training pipeline, covers some "LLM Psychology", and offers a few best practices:

https://t.co/HDJix905Gy

[Simon Willison @simonw]

I'm quite pleased with my latest attempt at a quick illustration of how prompt injection works

[Ilya Sutskever @ilyasut]

One of my hopes for the good AGI future is that people will become kinder to each other

[Andrej Karpathy @karpathy]

Good example of us not seeing max GPT-4 capability yet, imo. Prompt design, tool use, meta cognition strategies (eg idea of attempt, critique, retry, capabilities model, etc) are very likely to go a long way.

[Andrej Karpathy @karpathy]

This is a baby GPT with two tokens 0/1 and context length of 3, viewing it as a finite state markov chain. It was trained on the sequence "111101111011110" for 50 iterations. The parameters and the architecture of the Transformer modifies the probabilities on the arrows.

E.g. we can see that:
- state 101 deterministically transitions to 011 in the training data, so the probability of that transition becomes higher (79%). Not near 100% because we only did 50 steps of optimization.
- state 111 goes to 111 and 110 with 50% probability each, which the model almost learns (45%, 55%).
- states like 000 are never encountered during training, but have relatively sharp transition probabilities, e.g. 73% of going to 001. This is a consequence of inductive biases in the Transformer. One might imagine wanting this to be 50%, except in a real deployment almost every input sequence is unique, not present in the training data verbatim.

Not really sure where I was going with this :D, I think it's interesting to train/study tiny GPTs because it becomes tractable to visualize and get an intuitive sense of the entire dynamical system. Play with here: https://t.co/8jdceMLpqy

[Simon Willison @simonw]

Web LLM runs the vicuna-7b Large Language Model entirely in your browser, and it’s very impressive
https://t.co/mBf3Chgjn0

[Aman Hambleton @amanhambleton]

I just completed a speedrun featuring the King’s Indian Defense from both colours. I go into detail explaining common setups, plans, and what pawn structures you should aim for. 📈

All episodes from 800 ELO up to 2000 ELO can be found here! https://t.co/l5E7kXYKsK

[vitomd @vitomd]

@KellensBram Thanks. Unfortunley no, is not longer actively developed

[vitomd @vitomd]

In the weekend I migrated my site and projects to a new server, if you find some error or wrong redirects let me know. These should work including https:
https://t.co/CPUusFJRgY
https://t.co/2nxlVgFQyb
https://t.co/Oq2Enx3OiP
https://t.co/aQOb3ASH8Q
https://t.co/gedPITTTta

[vitomd @vitomd]

@MrScaryMuffin @Anna_Chess It´s means there is a tactic, the number is the centipawns that you can gain.

[vitomd @vitomd]

My #chess projects: Hand and Brain Chess https://t.co/BYliMphgdj The BRAIN will highlight the pieces to move. You must play the correct move as the HAND. Inspired by some fun videos from @Anna_Chess
More chess projects at https://t.co/OsSUQQGCvW

[vitomd @vitomd]

@davidsover thanks for let me know, should be working now. The https certificate expired some days ago.

[lichess.org @lichess]

Of course, less time=more blunders, but is it possible that the warning sound you hear on Lichess when you are low on time causes you to blunder? There is a noticeable bump at that time in different time controls. Full article: https://t.co/engcTndkqY

[vitomd @vitomd]

New Gem! Easy way to integrate Sign in with Apple to a @rails application https://t.co/kSRp0yhHDZ
#apple #rails

[vitomd @vitomd]

I had my reservations with @docker and now I have 4 containers in my #raspberrypi consuming really low resources. It's easy to try new software, copy composer file, docker-compose up, if you don´t like it, stop and delete. I have to say that is really useful for this use case

[vitomd @vitomd]

If you want to #selfhost this is an excellent resource https://t.co/RHA3JfybeK there are alternatives to everything: Your own google drive, password manager, project tracking, finance tracker, bookmark manager, dashboards, etc

[vitomd @vitomd]

My #raspberrypi is awesome, I #selfhost a lot of services 24/7 without problems. For example: Pi-Hole, 3 Cams with Motion, a Node server, HomeAssistant, rpi-monitor, organizr and a 1tb usb driver (with samba server so it´s works like a NAS) and the cpu load is less than 1%

[vitomd @vitomd]

If you want to organize your notes and ideas I highly recommend Trilium https://t.co/ql157gAPwI . I tried several options (from Google docs to my own solution https://t.co/8LSm2FyoOh) but Trilium wins

[vitomd @vitomd]

This is a interesting approach to read (not fiction) books and I must say it's pretty good. https://t.co/X2mLJVE4nR

[Santiago Bartesaghi @santib6_]

Cool, but something even better is coming thanks to @brunvez 😉 https://t.co/OcGLsyVMWz

#ruby #rails #opensource

[vitomd @vitomd]

@santib6_ @brunvez Hope the PR get merged