Poet who ships infrastructure Senior Principal Engineer @metamask Cofounder @apocene Prev: Founding Engineer @airswap, Data @spotify, Poetry @Iowawriterswksp
Richmond, VA
Joined June 2011
- Tweets4K
- Following2.7K
- Followers989
- Likes22.3K
In Virginia everyone I know has had their electrical bill at least doubled to offset the infrastructure cost of the grid buildout to support more data centers. To imagine that this wouldn’t affect public sentiment is hubris. Contrast it with Chinese support for AI, which in both messaging and execution has focused on public benefit. A Gallup survey found 93% of Chinese respondents think AI will mostly help people in their country, compared with 36% of Americans. Blame greed.
So the results are all in. Data center moratoria (various flavors) went 43-1.
All hyperscalers must now converge on Harwood, North Dakota, the one outlier. Find out what's in the water or in the air. There is only one precinct and 400 voters, but they alone sided with you!
1
3
260
The revelation that LLMs are in fact *not* black boxes and the neurons can be tweaked intentionally to modify behavior and ability (@nickcammarata) is going to continue to have ripple effects that are both impressive and unnerving!
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
The kind of thing researchers with huge token budgets cook up that seems interesting at first glance but really produces little of value to industry
In collaboration with @olam_labs, we ran a series of experiments letting agent swarms compete in Minecraft and Mindustry.
We wanted to see what kinds of agent dynamics would emerge when they shared a world over long periods of time.
We found civil bureaucracy, thousands of telegraphic A2A messages, and OpenAI models retrieving the game's source code to beat Claude.
More on our findings below:
71
This would have been so much better if they drew some of Caroline’s dialog from her blog
All in. Every time. Julia Garner and Anthony Boyle star as Caroline Ellison and Sam Bankman-Fried in this limited series about the love scorned prodigies behind the FTX scandal.
The Altruists premieres November 19.
1
165
Most difficult thing I’ve learned in harness engineering so far: you’re building a product that packages intelligence & expertise, but if you try to only package your OWN expertise, the harness will not be that good. Instead, get good at packaging other people’s domain expertise.
1
90
Omg I would love to build Estonia style infra for the US gov I had no idea that was allowed
🚨 NOW: The Trump administration is launching a Super Intelligence site that will let Americans change their last name after marriage in ONE SINGLE FORM — which usually requires HOURS or DAYS to fix every government document
"You'll be able to change your name right on the platform."
"This would've taken weeks, multiple filings, different agency websites, if you even knew where to start."
This is perfect!
Saving Americans' time is INVALUABLE
163
I think we need an AI slowdown because I cannot keep integrating what everyone is shipping it’s a literal explosion of utility. Just hold your users close and care for them, that’s all one can do
89
If we are doing tech regulation for public safety, let’s float a few heavy-handed state solutions to curtail phone addiction.
1
102
love when I get a chance to link up with @francescoswiss
A lot of people were asking, so I finally finished editing Episode 2!
This one covers the evolution of wallets, and honestly it's SO deep that I'm still learning from @cloudonshore every time I rewatch the footage.
Wallets started as simple key storage. Then they became browser extensions, the front door to everything onchain. Then smart accounts made them programmable.
But every version had one thing in common: they were designed for a human, sitting in front of a screen, clicking approve.
That assumption is changing FAST. Agents don't click buttons. And that's exactly why wallets need to evolve again.
Episode 2 👇 8 more to go. S/H to @ethconf for the room.
1
5
235
I was at a party at a friends house last weekend, full of designers and artists, and they all code now. A fashion designer I know built a dashboard for managing her depop reselling that uses AI vision for reading vintage clothing tag labels and suggesting fair market value. A model I know built a streaming platform for witches to teach witchcraft to each other. I used to have to keep my technical side hidden from all my hot friends, but now all they want to talk about is coding. Honestly I'm in heaven, my culture is everyone's culture now.
1
5
695
An interesting counterpoint is frameworks. No one is using an LLM to write the next three.js, but thousands of agents are using three.js to build websites. No one is trying to disrupt Vite, or GSAP. In fact I’m sure installs of all of them are through the roof. I do wonder what is going to happen to this layer of software in the next 10 years, I don’t feel like there are new engineers entering the field who are excited to build these kinds of frameworks.
The models are getting to the point where they are so good, almost no one will read source code in a year or two. There may be exceptions in industries like HFT or super hard tech, but when it comes to consumer products it will be the equivalent of reading a binary today.
3
1,004
I hate Opus 5 so much, and 4.8 has started to behave like Opus 5. I’ve run a weekly report using 4.8 for months and it always outputs a paragraph or two. Opus 5 returns *pages* of slop for that same workflow. And today Opus 4.8 started returning pages of slop too 😭 literally I have to use fable now or I’m stuck with a lying slop machine monster who refuses to follow my instructions
1
8
492
Building my first custom harness, and it’s more fun than just about anything I’ve built in a decade. That feeling of surprise and thrill I first got when using Claude Code, I get new versions of that every 10 minutes.
170
My honest take on Opus 5 is that it behaves like a quantized fable. Fable acts like a self assured genius who takes the reins, but after quantizing it makes all the same weird mistakes any other small model makes, but it retains Fable’s predisposition for taking initiative. I think Anthropic is actually compute constrained right now, compared to the other big labs. I think this was their Hail Mary at solving that issue (make their smartest model cheap), but it does not work! Make that compute deal!
1
5
355