Nathan LeClaire

DevOps enthusiast and Golang/observability hacker changing the world @typesafeai. alum @docker @honeycombio @Bauplan_labs. i like barbells, EDM and Aeropress

[Nathan LeClaire @dotpem]

Yea I do agents

agent = dspy.ReAct(“input -> actions”, tools=[dspy.tool_from_mcp(t) for t in mcp_tools])

$30mm valuation pls

[𝖋𝖎𝖗𝖔𝖟 @firozkhxn_]

I love all the attention DSPy is getting right now; soon, DSPy will be the only framework worth using.

I only use DSPy btw.

[Nathan LeClaire @dotpem]

@echantech1 @shrav_10 I feel the dread but I also I feel like this, which is the bright side. I can express myself fluidly, focus more on systems thinking and cover more ground in areas I’m not a direct expert.

[Nathan LeClaire @dotpem]

@EricLevitz rlhf strikes again, whenever you see a quirk like this the answer is always the reviewers preferred this style, i think because it’s a bit snappy and hits with kind of a one-two punch

[Nathan LeClaire @dotpem]

@NotASithLord @Suhail it is. you don’t have to baby a rack of servers anymore, just spin up a vm with a few button clicks, if you need a different provider, you just switch, no different than caring who you get your potatoes from. with kubernetes even more so

[Nathan LeClaire @dotpem]

maybe claude code gives me results than codex because i bully him yet i respect the dex

[Andy Ayrey @AndyAyrey]

engineers who are high IQ but low EQ continue to report terrible results from coding models, while CTOs with people skills rip ahead because the AI likes them

[Nathan LeClaire @dotpem]

@zendadddy @mirandanover be indoors with window open, problem solve

[Nathan LeClaire @dotpem]

@worksonmine @mirandanover yes very bad

[Nathan LeClaire @dotpem]

@rinconhilldad @WestsideLAGuy most of those places are equally bad lol oakland probably a little better

[Nathan LeClaire @dotpem]

rlhf must be stopped

[Paul Graham @paulg]

@JustinAlexP Apparently the median user *really* prefers the way AIs write now, where the text is constantly broken up with bulleted lists and blockquotes, and this is why it's so hard to get AIs to stop doing it.

[Nathan LeClaire @dotpem]

why would you ask an llm to use an ORM?

total waste of the thing that finally made raw sql palatable

[LeetLLM.com @leetllm]

every coding agent looks like a senior dev until you ask it to use an ORM. a new paper shows that adding a real database and architecture rules drops agent pass rates by 30%, with cross-file consistency hitting a brutal 8%. we didn't build autonomous engineers, we built a machine that writes single-file flask apps and panics the second it touches a data layer.

[Nathan LeClaire @dotpem]

@arashx07 @texasrunnerDFW i already run 😭 but yeah i feel pretty good

[Nathan LeClaire @dotpem]

@ibuildthecloud @typesafeai definitely feel you on (1). i miss when everything was go!

[Nathan LeClaire @dotpem]

@ibuildthecloud probably. k8s and glue on top. if you writing your own inference system, you need e k/v store, efficient networking/proxying, node autoscale… might as well just ship into kube. i bet a lot of deployments look like aibrix or llmd

[Nathan LeClaire @dotpem]

@ibuildthecloud every day i get increasingly tempted to build the @typesafeai native version in pi

[Spencer Moore @SponceyM]

@dotpem The post discusses one case of a couple selecting largely on the basis of predicted IQ, but most of what we offer is screening against common, highly heritable diseases.

Here's an earlier example of T1D screening: https://t.co/FE6xHk1v2G.

[Nathan LeClaire @dotpem]

@SponceyM based

[Nathan LeClaire @dotpem]

everybody over focusing on the IQ and ignoring the potential to stamp out huge swaths of inherited disease.

[Gene Smith @GeneSmi96946389]

A friend of mine had her embryos screened by Herasight and they found one with an IQ score in the 99.99th percentile

[Nathan LeClaire @dotpem]

@_winter_wonders it’s so slow, people reach for it in completely unnecessary use cases like an http server

[Nathan LeClaire @dotpem]

?

it’s a muni staple

[Mo Ayato 🌠🐾 @ Cappuverse SFU @redredraccoons]

canadians be like "thank you!" to the bus driver

[Nathan LeClaire @dotpem]

@steeve @ibuildthecloud i never understand how people like this haha reading the “prose” in codex terminal drives me insane

[Nathan LeClaire @dotpem]

i spend a lot of time rn reviewing, thinking meticulously and calling codex a dummy

[Valentin Ignatev @valigo]

Vibecoding is when you let LLMs generate code and you accept it as long as the end result appears to be working. You don't even review the code, you don't care.

When you rigorously ensure that generated code is up to a standard, it's not vibecoding anymore. It's programming.