Kyle Mathews

Big announcement! I'm now a co-founder at @electricsql along with the @thruflo / @balegas and the brilliant team they've built. We're solving local caching of Postgres data. Electric is a VC-backed devtools OSS startup. We're close to our first beta release so check us out!

Models got really good at cad in the last year. I tested them a year ago and they were barely adequate and now I've trivially designed a bunch of parts. RL in easily verifiable domains go brrrr

This never gets old

Reply with a dataviz/chart for any of the following you've either built or seen: - The prettiest (regardless of accuracy or usefulness) - The smoothest(animation/interaction/etc) - The most complex - The most useful - Or whatever you want, just write it in

@gill_kyle yeah PR bodies go into the git commit history too so you're largely writing (or having the AI writes as it is) a through explanation of the PR. I don't think most of it is that interesting to human reviewers

Maybe the wrong take but I think much of the AI authored content is useful as a record of context or benchmarks I've asked it to run which I think help my coworkers AI/review process. I read through the whole thing and edit, but I also don't want to throw away good context 🤷

@tomcritchlow I can dig it up later. Traveling today

@kylemathews ooh interesting. Will check this out Do you have a full output from this process?

5.6 Sol is a beast at essay writing and I'll prove it. I fed ChatGPT my Almanac essay from yesterday and asked it to write a brand new essay.... Follow the thread for a link to the full chatgpt convo and the final essay and judge for yourself!

@tomcritchlow github.com/KyleAMathews/d… Yeah good writing is good writing

@kylemathews Interesting! can you share? I think sometime this year we're going to move away from pangram scores of "did AI write this" to "do the ideas have AI bias in them" Feels like a hard problem to crack tho?

@tomcritchlow I made a skill that takes a thesis and asks the agent to write an outline to help avoid precommidified ideas. It can be deflating haha as all the cool ideas you've thought of get reproduced precisely

That said - the most insidious thing here is not the words themselves but rather the concepts being used. LLMs tend to get fascinated by a handful of topics and ideas and it's almost impossible to spot... If we all used ChatGPT to write essays about management theory, how many would converge on coase/scott/cybersyn? This is the pre-commodification of ideas that @sachinnbenny talks about here: summerlightning.substack.com/p/llms-pre-com…

GPT-5.4 full at xhigh scored 51, exactly where Luna max sits today. GPT-5.4 costs $2.50/$15; Luna now costs $0.20/$1.20. In other words, roughly four months later, OpenAI is selling March’s full flagship intelligence at about one-thirteenth the token price.

In light of Gianni Infantino’s poor judgement this week, I am beginning to question whether Donald Trump was a worthy recipient of the FIFA Peace Prize after all.

spiritually, it feels inevitable that AGI is born in California as a uneconomic carnival gizmo, and then dragged to New York, caged and put to work by men in grey tall towers that poke and prod it until it spills trillions out filtering through PDFs

The deep truth on JD Vance is spoken by Noah Smith here: “I will never forget the "Haitians eating cats" lie. I will never let it go. It was so deliberate and monstrous and unprovoked and cruel, and no one has ever apologized for it…. JD Vance even admitted that they lied... 1/

@gill_kyle ahem!

just recreated basically every chart from Particl in like 20 min 😂

TanStack charts looks VERY good. Fixes lots of quirks of other libraries - framework agnostic (👋 Recharts) - d3 marks/scales instead of a single data key (👋 Chart.js) - inferred (!) types (👋 Observable plot) - headless/themeable (👋 Nivo) - primitives not components (👋 Visx)

@gwenshap @martin_casado Kinda depends what you mean by harness. If a very thin harness is persisting data somewhere that future runs can read (e.g. a codebase) then learning is happening

@martin_casado All my experiments tell me that (1) is not true. Models without harness don't "learn" from failures or feedback. Instructions and memory seem important. Post-training for my codebase and my taste may be better, but seems like way more overhead than a harness.

On harnesses, I vacillate between three beliefs: - the less harness, the better. Models are the magic - post training a model and harness is dramatically better and the model providers win - harnesses have real independent value from the model I have no idea which is right.

Ironically the biggest harm on the horizon is a depression caused by potential overbuild on AI infra

The most remarkable thing about AI safety discourse is the incredibly high ratio of discourse to actual harm — as major new technologies go, LLMs have been enormously and shockingly positive.

@0xblacklight @badlogicgames It's entered it's "I'll sound smarter if I use big words" phase

@badlogicgames it deadass threw the word "epsilon" at me out of NOWHERE "delta"? fine. "sigma"? ok sure bud "episolon" wtf dude

fable thinks i'm a galaxy brain multi-phd. i can't take this anymore. it's like it speaks in code. useless for design work.

@MikaelLirbank yeah keeping it as a coder is fine but yeah, completely unhinged to discuss things with it

@kylemathews reddit.com/r/ClaudeAI/s/Z…

I'm finding Claude 5 Opus pretty annoying to talk to. It makes a ton of assumptions so I'm constantly correcting it. I'm strongly starting to prefer LLMs that do nothing until I ask them.

Software engineering is becoming a hierarchy of loops. So today we're launching Mastra Factory: a system for agents to take software from issue into production. Get started with 𝚗𝚙𝚖 𝚌𝚛𝚎𝚊𝚝𝚎 𝚏𝚊𝚌𝚝𝚘𝚛𝚢: