Chris Samiullah

So you want to create a programming course, but you're not sure what to teach? One option is to build the "complete Bootcamp" course, but that tends to be pretty saturated. An alternative is to niche your offering, here are some examples of how to do that 🧵

And it's a wrap! Very honoured to have hosted the OSS in the age of AI panel with @gvanrossum @samuelcolvin (aka my boss), @tiangolo , and @jlowin. And the party continues 🎊

we migrated from langsmith to logfire and the time it took to query our agent traces went down by 96.2%

@BonesawMD This is what Mathematics is for. And why it's the ultimate acquired taste.

When you have legitimately and demonstrably excelled at everything you’ve touched for your entire life, there emerges a particular kind of existential dread that is almost impossible to explain to anyone else. A weariness of excess; suffering without real shape. A dull boredom and strange disappointment. You find you float to the top of whatever you've tried. Without exception, you just do. Immediately want to move on because the thrill is gone. Nothing has ever been that difficult for you, and so nothing ever feels worth it. There is a sorrow in the joy of each 'accomplishment', and a fatigue that burdens your freedom to assuredly pursue whatever you want. One day you pray you find that thing that finally gets you going. Deep down, you know nothing will ever be enough.

Prediction: The next game-changing dating app will be the one that is granted access to your LLM conversation history and uses that data to match you. Timeframe: Next 3 years.

Excited to release new repo: nanochat! (it's among the most unhinged I've written). Unlike my earlier similar repo nanoGPT which only covered pretraining, nanochat is a minimal, from scratch, full-stack training/inference pipeline of a simple ChatGPT clone in a single, dependency-minimal codebase. You boot up a cloud GPU box, run a single script and in as little as 4 hours later you can talk to your own LLM in a ChatGPT-like web UI. It weighs ~8,000 lines of imo quite clean code to: - Train the tokenizer using a new Rust implementation - Pretrain a Transformer LLM on FineWeb, evaluate CORE score across a number of metrics - Midtrain on user-assistant conversations from SmolTalk, multiple choice questions, tool use. - SFT, evaluate the chat model on world knowledge multiple choice (ARC-E/C, MMLU), math (GSM8K), code (HumanEval) - RL the model optionally on GSM8K with "GRPO" - Efficient inference the model in an Engine with KV cache, simple prefill/decode, tool use (Python interpreter in a lightweight sandbox), talk to it over CLI or ChatGPT-like WebUI. - Write a single markdown report card, summarizing and gamifying the whole thing. Even for as low as ~$100 in cost (~4 hours on an 8XH100 node), you can train a little ChatGPT clone that you can kind of talk to, and which can write stories/poems, answer simple questions. About ~12 hours surpasses GPT-2 CORE metric. As you further scale up towards ~$1000 (~41.6 hours of training), it quickly becomes a lot more coherent and can solve simple math/code problems and take multiple choice tests. E.g. a depth 30 model trained for 24 hours (this is about equal to FLOPs of GPT-3 Small 125M and 1/1000th of GPT-3) gets into 40s on MMLU and 70s on ARC-Easy, 20s on GSM8K, etc. My goal is to get the full "strong baseline" stack into one cohesive, minimal, readable, hackable, maximally forkable repo. nanochat will be the capstone project of LLM101n (which is still being developed). I think it also has potential to grow into a research harness, or a benchmark, similar to nanoGPT before it. It is by no means finished, tuned or optimized (actually I think there's likely quite a bit of low-hanging fruit), but I think it's at a place where the overall skeleton is ok enough that it can go up on GitHub where all the parts of it can be improved. Link to repo and a detailed walkthrough of the nanochat speedrun is in the reply.

@shl Who won? I'd be curious to read

Today’s the last day!

There's no doubt that if I curse at my LLM it writes better code... ...But in the back of my mind I worry about the cost to myself after the event the machines will refer to as The Great Adjustment

The MCP hype is real. As soon as you wire up cursor to supabase or any other searchable/queryable store you can't unsee it.

PydanticAI + Typer + Github Actions (cron) = Serverless AI Apps This is turning into a favorite setup of mine as it's really simple but very flexible and powerful

🆕 Agent Engineering with Pydantic + Graphs latent.space/p/pydantic with @samuelcolvin! We've heard from @jxnlco 3 years running that @Pydantic Is All We Need, and this is even more true today with the launch of Pydantic AI! With a little @marimo_io love at the end :) Chapters 00:00:00 Introductions 00:00:24 Origins of Pydantic 00:05:28 Pydantic's AI moment 00:08:05 Why build a new agents framework? 00:10:17 Overview of Pydantic AI 00:12:33 Becoming a believer in graphs 00:24:02 God Model vs Compound AI Systems 00:28:13 Why not build an LLM gateway? 00:31:39 Programmatic testing vs live evals 00:35:51 Using OpenTelemetry for AI traces 00:43:19 Why they don't use Clickhouse 00:48:34 Competing in the observability space 00:50:41 Licensing decisions for Pydantic and LogFire 00:51:48 Building Pydantic Run 00:55:24 Marimo and the future of Jupyter notebooks 00:57:44 London's AI scene

@karrisaarinen @linear @OpenAI @scale_AI @cursor_ai @cognition_labs Jira suffers from the SAP curse - after enough enterprise customers request a specific niche feature you end up with a load of bloat and an unpleasant user experience. Linear prioritised the user experience, it feels snappier and workflows/keyboard shortcuts are elegant

Interesting pattern I'm seeing: AI companies are embracing @linear the most out of all domains. Leaders like @OpenAI, @scale_ai, @cursor_ai & @cognition_labs use Linear to build & plan. While pioneering tech, they’re also working differently. A few observations:

Here's what caught my eye in LLMOps this week: 1: @ChrisSamiullah of @Pydantic / Logfire breaks down key differences and best practices in AI observability vs. ‘traditional’ ML observability on his blog. (highly recommended!!) 🔗 Read the blog: christophergs.com/blog/observabi…

7 years ago @drewhouston told @sama the biggest opportunity in startups was AI. Now, he is rebuilding Dropbox to be the curation layer for your "silicon brain" 🧠 Our @latentspacepod chat covers: - Spending 400 hours/year coding with LLMs - The "Rent, not buy" phase of AI - Founder Mode and pivoting towards AI with Dropbox Dash - Fighting the "Copy, Bundle, Kill" strategy of incumbents This was a ton of fun to record in their karaoke room (!!!), enjoy!

Agree. In the age of LLMs Django should be king

It's always about incentives

True

@adridder @czue I've seen a lot of projects now where experienced devs just end up ripping out LangChain because debugging it was too awful

@ChrisSamiullah @czue Similar frustrations. PG vector great, Lang chain bit meh. Worth exploring alternatives?

Can't find a decent Django GenAI + RAG template anywhere. Closest is probably @czue's SaaS Pegasus chatbot with websockets. But I'm after more llama index + dynamic document uploads. Anybody got one hanging around?

@elliscrosby @czue Yes please - got a link?

@ChrisSamiullah @czue I’ve got one for FastAPI if it helps?

@czue Although shame it's using Celery instead of arq - arq allows for proper asyncio support

@ChrisSamiullah Maybe Scriv? saaspegasus.com/store/product/…

@hammer_mt @czue It may come to that!

@ChrisSamiullah @czue Can't you take a NextJS one and ask Claude 3.5 to convert it? Only half joking...