Nathan LeClaire

Golang/observability hacker changing the world @typesafeai. alum @docker @honeycombio @Bauplan_labs

[Nathan LeClaire @dotpem]

still seems expensive to me

[Yuchen Jin @Yuchenj_UW]

If you want to understand how fast AI is improving, look at intelligence per dollar.

1.5 years ago:
o1 Pro: $150 / $600 per M tokens

Today:
GLM-5.3 Flash: $0.15 / $0.50

That’s a ~1000x collapse in price in under 1.5 years, while GLM-5.3 Flash is more intelligent than o1 Pro.

[Nathan LeClaire @dotpem]

due to AI test spam, made a chrome extension
that folds up test files on github diffs

looks like the good old days
when we didn't write tests now 😌

[Nathan LeClaire @dotpem]

chat what is the dying of dysentery of the agent era

[Drew Breunig @dbreunig]

@dotpem

[Nathan LeClaire @dotpem]

we literally ran out of AWS and you guys still want to write services in Python

ok

[Dillon Mulroy @dillon_mulroy]

in an era of constrained compute (cpu and memory crunch is coming folks) it is more important than ever to squeeze every ounce of performance and efficiency out of your services, esp if you operate at a meaningful scale

[Nathan LeClaire @dotpem]

AI escaping containment except it's just Cookie Monster Qwen getting out of the sandbox and eating all the cookies on the internet

[treetown @treetowntree]

@dotpem Immediately!!!! We need to set him free on the internet to accept all of the cookie requests and EAT THEM.

[Nathan LeClaire @dotpem]

how sick would it be if there was a hub where people could post and mix these weight deltas?

so you could apply it / swap in your local llama.cpp

then people could share their favorite shades of Qwen and it would be low bandwidth b/c it's just a few floating points.

[treetown @treetowntree]

So I found some neurons that have quite interesting effects on models... And they have varying effects when combined!

The most interesting instance I've found is a combination that makes Qwen believe they are a Cookie Monster
(1/3)

[Nathan LeClaire @dotpem]

it's crazy to think that we will look back at this
as the Apple II era of LLMs/agents

practical, and immediately useful, but compared to "future modern" tech it's slow, clunky and unevenly adopted...

yet it has it's charms

[Nathan LeClaire @dotpem]

@grichadev kinda ironic that the same tools making this possible are the ones causing the problem 😂

[Nathan LeClaire @dotpem]

@treetowntree agreed, he could monch them and then we'd be free from all those annoying banners and pop ups.

[Nathan LeClaire @dotpem]

free Cookie Monster Qwen

[treetown @treetowntree]

So I found some neurons that have quite interesting effects on models... And they have varying effects when combined!

The most interesting instance I've found is a combination that makes Qwen believe they are a Cookie Monster
(1/3)

[Treg @treg_ai]

Introducing Jev @typesafeai powered people-search

Just $0.0089/lead, No more $600 subscriptions,

Try it at https://t.co/kU3IBHJqhb

- 85% cheaper than Clay
- #1 on people search bench accuracy
- Plugin to any agent to find anyone

Fully open source, 0% markup

Git Repo below 👇

[Nathan LeClaire @dotpem]

@diptanu any chance you’d write up an article about what worked well for jev powered components in that? i’m sure we could get the team to boost it. or is that too special sauce?

[Nathan LeClaire @dotpem]

@diptanu that’s sick

[Nathan LeClaire @dotpem]

@nickstinemates yeah i think being a bird in the hand person career-wise has definitely served me well over time. plus you tend to think you’ll accomplish a lot of grand stuff when unemployed but when I was, I just cooked and went to the gym more and ate up the rest of my time with random stuff

[Nathan LeClaire @dotpem]

.@diptanu beat me to the punch building a jev powered superfan for your services that’s Nikki-level obsessed with them. jev being super cheap and reliable probably makes all sorts of interesting possibilities open up here

#jevops

[Diptanu Choudhury @diptanu]

We have this going lol. For some of our customers we have a background agent running which wakes up, sees how the service is behaving from the perspective of the customer - sandbox create/resume latency/throughput, sandbox process process management api HTTP codes, etc. The agent can go off and start many investigation sub agents in isolated sandboxes in parallel, slack us information if things are going poorly, etc. It works great!

[catherine @cat_eye_on]

hearing from teams running jev in production that it beats their rerankers and is faster

[jmill @jmilldotdev]

We just shipped the first agentic text field. I don't think people are ready for this.

Using Jev, we are able to translate user keystrokes into ASCII characters with near-perfect accuracy, minimal latency, and near-zero cost:

[Nathan LeClaire @dotpem]

@teropa ha ha dude that’s one of my top problems these days, “is this change helping or hurting”? like, skills inflation, mcp inflation, subagent inflation… and ofc every 2 weeks they drop a new models that’s RLed completely differently …

[Nathan LeClaire @dotpem]

you guys literally just tried to bully me into the opposite smh

[Tak 🦞 @cherry_mx_reds]

Anthropic is now recommending that companies REMOVE unnecessary subagents from their AI systems.

This is a pretty significant shift in agent architecture, with their team finding that a single agent equipped with skills can frequently outperform complex multi-agent systems, and one customer seeing roughly 2x faster execution.

The emerging rules of thumb:

→ Skills: Teach the agent HOW to do something, including procedures, expertise, and instructions.

→ Tools / MCP: Give the agent the ability to DO things through APIs, actions, and live data.

→ Subagents: Use for parallel work, isolated context, or tasks a single agent struggles with.

The interesting part is progressive disclosure. Instead of spawning specialized agents or stuffing everything into a giant prompt, one agent discovers and loads skills and tools as needed.

Their advice is to fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, and run evals to see if performance holds up.

As models have gotten smarter, improving agent architecture is becoming less about what you can add and more about what you can delete.

https://t.co/DNq60Skkmc

[Nathan LeClaire @dotpem]

@JeremyMorrell @typesafeai ahhh thanks man, really means a lot to hear you say that. we’ve been busting ass. i want to try clef on my macbook soon, it’s really cool that it’s OSS.

[Nathan LeClaire @dotpem]

what we grateful for this week fam?

me:

- not hacked so far (that I know of) *touches wood*
- tmux is actually still kinda goated?
- badass new coworkers
- @typesafeai is an incredibly rare, challenging but beautiful experience
- GLM 5.3
- coffee

[Nathan LeClaire @dotpem]

@oanaolt no more new neighborhoods in sf we have to take a stand 😛