Kayvane Shakerifar retweeted
Watching 4-5 Codex computer use agents work on separate tasks on my computer right now is pretty close to AGI for me personally.
And I get to stay focused on what I need to do.
The recent speed gains the team has cooked up are killer. Give it a try if you haven't yet!
Kayvane Shakerifar retweeted
major price cuts today:
*80% drop for GPT-5.6 Luna, now $0.20 per million input tokens and $1.20 per million output
*20% drop for GPT-5.6 Terra, to $2/$12
*GPT-5.6 Sol gets Fast mode in the API, up to 2.5x the speed for 2x the price, same intelligence
Kayvane Shakerifar retweeted
More of Codex is rolling out across Europe this week.
We’re bringing Computer use, the Codex Chrome extension, personalized memory, and Chronicle to Codex users in the EEA, UK, and Switzerland.
developers.openai.com/codex/…
Kayvane Shakerifar retweeted
A talk I gave a few weeks ago.
Software fundamentals matter more than ever. Here's why:
youtube.com/watch?v=v4F1gFy-…
Kayvane Shakerifar retweeted
Replying to @MohapatraHemant
for chips, AI and data centers from first principles I would self recommend:
tensoreconomics.com/
Kayvane Shakerifar retweeted
This is what I worry Europe will get negatively polarized into: an ideology taking pride in a neat, sanitized online environment free of evil corporate and fascist pathogens.
I hope European govs do not go this way, and instead take a Pirate Party approach of user empowerment.
First, what's wrong with the tweet I'm quoting:
The idea that there should be "no space" for something you dislike is fundamentally a totalitarian and anti-pluralistic impulse. It's incompatible with being in an environment that you do not fully control.
This is especially true for categories that are subjective and controversial, because you end up trying to fully remove things you think are pathogens, when other people have good faith disagreements, and because you give yourself the maximalist goal of not even giving them breathing room, you create conflict and end up building the machinery of technocratic authoritarianism to impose your victory in the conflict.
So sorry, if you want to be a free society, you have to bite the bullet that some people, somewhere, will be selling things that you consider dangerous and saying things you consider disinformation and vicious lies.
What is the goal to shoot for?
You want to create an environment where those things don't dominate. This is the problem with twitter today: not that it's a safe space where 1000 people talk to each other in a corner about how heritage americans are the master race and putin is good or whatever, but that that crap gets shoved in our face on a mass scale, and the algorithms actively favor it.
The right metaphor is not castles and walls, but biological - think, why European forests don't have tropical lizards.
Having incentives for social media platforms to have less of those things instead of more is fundamentally reasonable, @audreyt has talked about how Taiwan has done something similar.
You also want to do this in a way where it's clear what the underlying principle is, so it's not a vehicle for imposing arbitrary and frequently changing expert-consensus agendas.
You also want to empower users, rather than working against them. People want to see and buy good things instead of bad things. Often the problem is that competition is too difficult in the current market. I actually supported the USB-C standardization mandate; it created more interoperability and thus improved competition and convenience. I would support incentivizing social platforms to be more open, and to be more transparent (eg. my proposal to require algorithms to be continuously published with a 1-2 year delay, with zk-proofs to ensure that the algorithm being used in real time exactly equals the one that gets published later)
Being able to better identify what messages are coming from what communities is also good, though I don't support the direction of banning anonymity of individual posters, rather I would want to see more macro-scale analytics, eg. seeing what communities are most strongly saying and amplifying content that semantically matches a particular idea; this can be done in privacy-preserving ways.
There is a real opportunity to reaffirm freedom of speech in a unique and different way, that emphasizes pluralism and pushes against unbalanced attempts to manipulate the discourse by individual powerful actors. We want to do this, not go down the dark path of having something that claims to support fundamental rights but actually is not trusted by anyone to be anything other than the fundamental right to follow the footsteps of a few technocratic experts.
𝗡𝗢 space for cyberbullying.
𝗡𝗢 space for dangerous products.
𝗡𝗢 space for hate speech.
𝗡𝗢 space for scams.
𝗬𝗘𝗦. With the Digital Services Act, what is illegal offline remains illegal online.
🔗 link.europa.eu/gbRf9h
Kayvane Shakerifar retweeted
Meet Tülu 3 -- a set of state-of-the-art instruct models with fully open data, eval code, and training algorithms.
We invented new methods for fine-tuning language models with RL and built upon best practices in the community to scale synthetic instruction and preference data.
Demo, GitHub, technical report, and models below 👇
Kayvane Shakerifar retweeted
I've been helping teams with prompting and was sad some didn't understand the basics even as they used advanced techniques and complex tools.
Thus, I wrote about the fundamentals and how to apply them effectively.
Qn: What have you found to work well?
eugeneyan.com/writing/prompt…
Kayvane Shakerifar retweeted
my game where you lure AI - powered NPCs to your alien ship
play now in the browser:
playsnatched.com/
Kayvane Shakerifar retweeted
Can AI rewrite our human genome? ⌨️🧬
Today, we announce the successful editing of DNA in human cells with gene editors fully designed with AI. Not only that, we've decided to freely release the molecules under the @ProfluentBio OpenCRISPR initiative.
Lots to unpack👇
Kayvane Shakerifar retweeted
Alongside our Mixtral 8x22B release, we are releasing our tokenizers, which go beyond the usual text <-> tokens, adding parsing of tools and structured conversation.
Repo: github.com/mistralai/mistral…
Guide: docs.mistral.ai/guides/token…
Kayvane Shakerifar retweeted
Today we’re highlighting the work of Kayvane Shakerifar, who created an open source Python wrapper for Brave Search API that helps to speed up the process of building with our independent search index.
Kayvane Shakerifar retweeted
As our journey with YC W24 wraps up, we're thrilled to share a special gift with you all! 🎁
We are releasing one of first model as part of @Ragas_io for synthetic test data generation - ragas critic LLM to replace GPT-4 as critic. ⭐️
▸ Finetuned + GPTQ quantised Qwen 1.8B model on synthetic data to reduce the cost of synthetic test generation by 90%. Can give upto 200 tokens/sec on single A10 instance.
Try it out here : docs.ragas.io/en/latest/inde…
Model : huggingface.co/explodinggrad…
Video tutorial explaining concepts and how to use: youtu.be/cOyUnGG321Q
We will be working on and releasing more of these types of models to solve the problem of LLM application evals and testing in coming months. OSS ❤️































