Tony retweeted
آندری کارپاتی، یکی از همبنیانگذاران OpenAI، تقریباً تمام چیزهایی رو که درباره ساخت هوش مصنوعی میدونه، رایگان روی GitHub منتشر کرده!
از ساخت GPT از صفر تا هوش مصنوعیای که خودش آزمایش میکنه، تحقیق میکنه و از نتایجش یاد میگیره.
۶ پروژه با مجموع بیش از ۳۰۰ هزار ستاره ⭐
۱. Autoresearch | 92.7K ⭐
هوش مصنوعیای که شب تا صبح خودش آزمایش انجام میده، نتایج رو بررسی میکنه و فقط راهکارهایی رو نگه میداره که واقعاً جواب دادهاند.
github.com/karpathy/autorese…
۲. nanoGPT | 64.4K ⭐
یاد بگیر چطور یک مدل GPT رو از صفر آموزش بدی؛ پروژهای که برای خیلیها نقطه ورود به دنیای LLM بود.
github.com/karpathy/nanoGPT
۳. nanochat | 60.2K ⭐
با حدود ۱۰۰ دلار، نسخه شخصی خودت از ChatGPT رو بساز.
github.com/karpathy/nanochat
۴. LLM Wiki | 50K ⭐
ایدهای ساده اما قدرتمند برای ساخت یک دانشنامه شخصی و قابل توسعه برای هوش مصنوعی؛ چیزی شبیه مغز دوم.
gist.github.com/karpathy/442…
۵. llm.c | 31.4K ⭐
آموزش یک مدل زبانی با زبان C؛ بدون پنهان کردن پیچیدگیهای فنی پشت فریمورکهای آماده.
github.com/karpathy/llm.c
۶. micrograd | 18.7K ⭐
در چند صد خط کد بفهم شبکههای عصبی چطور یاد میگیرن و گرادیانها چطور کار میکنن.
github.com/karpathy/microgra…
پیشنهاد من:
از micrograd شروع کن.
بعد برو سراغ nanoGPT و nanochat.
این مسیر کمک میکنه بهجای اینکه فقط با ابزارهای AI کار کنی، بفهمی پشت پرده واقعاً چه اتفاقی میافته.
تفاوت استفادهکننده از هوش مصنوعی با کسی که میتونه سیستمهای هوشمند بسازه، دقیقاً از همینجا شروع میشه.
این ۶ ریپازیتوری رو ذخیره کن؛ گنجینهای از دانش رایگاناند.
We’re renaming our Max plans:
Max 5x → Max
Max 20x → Ultra
Max is still $100/month, and Ultra is still $200. If you’re a subscriber, there’s nothing you need to do.
We’re also removing the 5-hour limit for verified Ultra accounts, starting today with most existing subscribers. Newer accounts keep a temporary limit until they’re verified, which helps us prevent fraud and abuse. You'll get an email when it's removed.
Weekly limits remain unchanged. View our plans here: claude.com/pricing
Tony retweeted
Philly Tech Pulse - October 31, 2026 Comcast’s nationwide overhaul will cut 634 jobs at its Philly headquarters technical.ly/workforce/comca… #CodeAndCoffeePhilly #PhillyTech #Startups
Tony retweeted
These pictures are 100% representative of how much time I spent laughing and smiling at the event last night.
It's so much fun to get out with other nerds. Find your local meetups or start one if there aren't any!
Tony retweeted
observations from adding swarm communication dynamics into my Eval/Environment Creation Factory
- Agent to Agent DM can create inefficient information silos compared to posting in Public Slack Channels -> many human orgs have discovered this
- Agents are pretty good at self organizing into niches and picking up work! They settle into personas but importantly will swap personas to pick up work if they don’t have anything to do. I didn’t ask for multiple task verifiers but they emerged
- Agents benefit from being able to procure more infra! Especially for testing, it’s hard to know what specs they need for rollouts, so they can specify and edit as needed
- Agents have their own infra, they can work on stuff privately in their own box and break anything
- Durability is important, many things break so a central controller can always resume a dead agent
- Swarms are wall clock much more efficient than single agent, this matters a lot for things that need human review. But they’re often marginally to a decent chunk less token efficient. I think this is a good tradeoff
- I’m fairly confident that agents are better than us at specifying roles for themselves, we need to focus on good goals
AI swarms are already showing promising (or dangerous) results in frontier labs.
But how to implement them in the open community? And how do they scale?
We built SwarmChat, an open-source API-based Slack for agents.
Then we burnt half a trillion tokens, ran a series of swarm-scaling experiments, and tested various swarm structures on ProgramBench to study this question!
Check out the blog here: swarmchat.io
We found a bunch of interesting results:
1. Swarms get the task done faster. For instance, 16-agent swarms reached the performance of a single agent 6x faster.
2. Swarm scaling is a third axis of scaling. The more agents, the better, and anyone can already start scaling with SwarmChat!
3. Different models communicate differently. GPT-6 is chatty and over-communicative, while Claude prefers to work in relative silence!
4. Different swarm structures barely have any benefit on aggregate. But on specific tasks, team structures give clear gains.
5. Bonus: SwarmChat has a Slack-like UI for us humans to monitor swarms! Makes monitoring a lot easier! We are already using SwarmChat for WIP projects, and we were able to catch cases of swarms breaking the 4th wall (sorry, more details are for a future post)
Code: github.com/Pranjal2041/swarm…
w/ @JangLawrenceK , @kohjingyu , @dan_fried, @rsalakhu, @wellecks, @Yngkvn , @nicoup
Tony retweeted
despite the unfortunate timing, we take disclosures seriously at @typesafeai
at approximately 6:49PM this Friday evening, we became aware of a massive unauthorized leak of ALL customer noul outputs
in effort to maximize transparency, here is the link: gist.github.com/diogo149/810…
Tony retweeted
Jev as a judge for LLM evals!
JevEval uses Jev as the judge for custom LLM evals instead of relying on another generative LLM for every evaluation.
The idea is pretty simple.
You define the exact decisions you want the evaluator to make, Jev evaluates each one and returns probabilities, and DeepEval uses those results to calculate the final score.
For example, with an AI agent you can evaluate the trace itself, checking whether the agent selected the right tools, followed the expected sequence of actions, and actually completed the task.
Jev handles those semantic judgments, while the scoring logic stays outside the model.
That is the main difference from a normal LLM-as-a-judge setup.
With a generative judge, the model often has to interpret the rubric, reason about the response, and generate the evaluation itself.
With JevEval, you define the decision structure upfront.
You can also assign different weights to different checks or use `strict_mode` when every requirement needs to pass.
So if one part of the agent workflow matters more than another, you can reflect that directly in the scoring logic instead of leaving the entire evaluation to the judge model.
Key capabilities:
- Evaluate agent outputs and traces with Jev
- Define custom checks as bounded decisions
- Use yes/no, score, and closed-choice questions
- Weight different checks or require all of them to pass
- Evaluate both single-turn outputs and multi-turn conversations
This is where Jev makes a lot of sense for evals.
If you already know what needs to be checked, you can define those decisions upfront, let Jev handle the semantic judgment, and keep the final scoring deterministic.
I’ve shared the DeepEval post in the comments.
Tony retweeted
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
🤖 Made with AI
Tony retweeted
don't sleep on @tanaysingh789, super talented, and just getting started
Tony retweeted
Stop manually making flowcharts. Try this GitHub skill with Claude Code.
I found an open-source tool that turns a plain description of a software system into a clean, interactive diagram.
It's called Archify, and it has more than 66,000 stars on GitHub.
You add it as a skill to an AI coding agent like Claude Code, Codex, Cursor, or OpenCode. Then you describe the system in the chat in one line.
Archify turns that line into a single HTML file you can open, click through, and send to anyone.
If you point your agent at an existing codebase, it reads the code and Archify maps how the pieces connect.
I liked five things about it.
1. It makes five kinds of diagrams, including system maps and step-by-step workflows.
2. You can click any piece and trace what connects to it.
3. It checks the layout before it hands the diagram over, so arrows don't pile up and labels don't overlap.
4. It has dark and light themes and exports to PNG, SVG, or a short video.
5. You keep editing in plain chat, with requests like "add Redis" or "highlight the rollback path."
You don't need a codebase to use it. A description in the chat is enough.
Tony retweeted
JEV is INSANE.
We gave it 400 companies and one candidate profile.
In 12 seconds, it predicted which jobs the candidate had the highest chance of getting, assigned a confidence score and detected job-candidate mismatches.
All for just $0.0005 It can also score companies, analyse your experience, match you with the right roles and identify the opportunities you’re most likely to get based on your profile.
Coming soon to @textbackdoor
Comment “JEV” for early access.
Tony retweeted
Playing with Jev adapter for Codex @CompleteSkeptic @sama @elonmusk - this should be the default. Current computer / browser use is a tad slow.
Lots of improvement here - from cutting the subagent handoff time to warm up time and improving the loop itself - but i think i got it to a working state.
Jev is an absolute unlock
Tony retweeted
github.com/githubnext/localj…
Not everyone on the team has access to Jev yet. Spent a morning cobbling together a poor man's Jev on top of omlx for local use. Benchmarked and eval'ed a variety of models including diffusiongemma and a variety of autoregressive models (Qwen MoE, Gemma 4 MoE, and Gemma 4 e4b/e2b.) Benchmark report is in the repo.
This is 100% promptcoding but hey the evals look okay, speeds are pretty good on local machine (m5 max 64gb), and the LocalJev server exposes an API that can be used with the normal Jev API wrapper libraries.
Tony retweeted
jev is insane 🤯
Here is Jev playing subway surfers at super human speed, and also playing 50 games at once.
cost less than a cent to do this run.
Jev does not replace llms like astra or fable, but opens up an entirely new world of capabilities.



































