grass is blue

Grass is blue

it’s been a while. Back to the land of blue :)

Boss man did it!! DO’s will be the agent infra for the future.

Super fun to speak at a packed ai engineer london. My talk was about how we eval agentic RAG with Langsmith Thanks @robbiehudson.bsky.social and crew for having me and for organising a great event 🙏

in 2025: - build a cursed programming language - publish my RAG stuff as real research - build more models with zml - generate 100x more tokens

learnt alot in 2024: ML -> transformers, finetuning data mindset AI -> search stuff, agent guardrails, evals cloudflare dev -> awesome modal labs -> containers in python Data engineering -> DAGs, lineage, access control, etl at scale Comp sci -> memory management, search algos, basics of zig

We have been building something for a while at StackOne with no idea what it should be called. Google just told us. We are building an extension an extension for the whole HR tech ecosystem.

remote work hack take lunch after a long meeting and go do some exercise. come back with a clear head, knowing what to do. Nobody cares what time you take lunch anyway.

i've been building a thing 🦓 the api is coming together. the model compiled by zml I think this leans well on zml pros. One model, many backends, zero code change.

so a bit like this

kinda enjoying this.. Leetle #8 1/6 12:31 [8 lines] 🟩🟩🟩🟩🟩🟩 leetle.app

Slow memory of the stack solution Leetle #7 1/6 6:32 [10 lines] 🟩🟩🟩🟩🟩🟩 leetle.app

Trying to define the AI Engineer role, I just explained it to a friend on 2d axis. On the x axis we have designer to FE to BE to ML to Data Science On the y axis we have cloud to hardware AI engineer blob sits right in the middle. wip but I think this is a nice way of thinking about it.

Streaming LLM tokens is a hype. Give it a year and peeps will just use rapid models for quick answers and the slower models for more complex work so much faff the last two years just for streamed tokens in chat. Horrible UX

London looking pretty this time of year

French have a word for being thrown from a window. défenestrer

EU ftw

made a lil website for my granny. She needed some photo storage for her (soon to be an online) bookshop. Build her a nice little admin panel and it has dark mode :)

Good year. pushed pretty hard and got a lot better

lfg

Making this thing I have wanted for the last year :)

python is so broken but the fact that this is a thing.. js ecosystem is no better wtf

2024 - moved in with my gf ❤️ - new job at StackOne working on LLMs 🤖 - started ADD with MC - 4 events, over 1k people registered - won two hackathons with legends 🏆 - learnt to paraglide - ran a decent half marathon (1:36) Next year I wanna read more books, do some research and climb a few v6s

Repeat after me: more AI chat here more AI chat here more AI chat here more AI chat here more AI chat here more AI chat here more AI chat here more AI chat here more AI chat here more AI chat here more AI chat here more AI chat here more AI chat here more AI chat here more AI chat here

Check out the latency increase by using gpt-4o instead of a specific model tag This has to be a bug?

Just noticed some mad inconsistencies with OpenAI model latency. Who can I tell?

The new Gemini models thinking mode is really good. Like really really good

I want angry reacts on LinkedIn

I look for light dev tools. That's it. Has to feel breezy

Deployed a brand new product to prod today. Low key stoked

If I ever stand trial I want O1 as my judge. Gemini would give me life in prison for not holding the door open for an elder but the slightest bit of contradictory evidence would have O1 singing my praises

Took me 20 minutes to work out how to get a Gemini API key smh. I had a billing account already setup

I knew gemini was the best LLM as a judge. Deepmind just proved me correct. www.kaggle.com/facts-leader...

the OCR-killer huggingface.co/collections/...

meta^2 prompts ftw

French is so nuts I’m sat next to two natives literally googling the pronunciation of « 4 eggs »

day 11 of Advent of ML (nice to be back) and we are talking scaling at test time. Do models reason? what is scaling at test time and will it lead to mythical AGI level reasoning? find out more about this new scaling law: mattzcarey.com/blog/advent-... #blogvent #adventofml

I'm a pretty big fan of how fast and cheap dynamodb is. but the devex = ropey. Literally writing entity shapes in comments not forget them. dynamodb-toolbox makes dev ex as pretty as working with best orms I KNOW is not an ORM but I'm smooth brained (@jeremydaly.com) github.com/dynamodb-too...

I just completed "Guard Gallivant" - Day 6 - Advent of Code 2024 #AdventOfCode adventofcode.com/2024/day/6 brute forced this so hard. Not safe or fun. Enums in zig are cool tho

Been living in London for almost 3 years, just got a water filter. Man it tastes like home I’m never going back to the sludge

went to a Christmas party last night with seemingly a bunch of musicians. Was awesome.

I love when people think they know best online and end up arguing with the dude that built the actual thing. never gets old

probs won’t do an advent of ML today. Need a day off. I have something cooking with ZML. High performance inference in Zig. More code examples for the next few days :)

day 7 of Advent of ML and we are talking Transformers. one of the greatest discoveries in the last decade and the paper didnt win a single award in NeurIPS at the time. find out more about the discovery which made LLMs possible: mattzcarey.com/blog/advent-... #blogvent #adventofml

When the world wakes up to the power of evals. Tracing services gotta invest in their compute. 🙏 for the SREs

So you've built something with AI, does it work? do you even know? Day 6 of Advent of ML is on evals and how you measure the success of your system pre and post deployment. Hope you enjoy :) mattzcarey.com/blog/advent-... #blogvent

Successful day. Turns out writing about best practises makes you more conscious of following your own advice. 60% still sucks but getting there.

It's already day 5 of Advent of ML (also known as #blogvent) 😇 Finishing up on the mini dive into retrieval today. If you are not using these you probably should.. reranker! An upgrade to @cohere.com rerank v3.5 bumped our internal retrieval evals 5%!! mattzcarey.com/blog/advent-...

go.bsky.app/VBXf2SB If anyone wants to join, feel free to write a comment! #indiehacker #buildinpublic