joy larkin (@joy) header
joy larkin (@joy) profile picture

Technologist. AI Marketing & Comms/GTM/Community Leader. Prev: Multiple startups-3x exits. I like robots & I'm excited for safe superintelligence. NJ/PHL✈SF ♥︎

SF Bay Area
Joined November 2006
Photo by joy larkin (@joy)Photo by joy larkin (@joy)Photo by joy larkin (@joy)Photo by joy larkin (@joy)Photo by joy larkin (@joy)Photo by joy larkin (@joy)Photo by joy larkin (@joy)Photo by joy larkin (@joy)Photo by joy larkin (@joy)Photo by joy larkin (@joy)
The riding in a robot car feeling hasn’t gotten any less wondrous yet.
6
4
39
32,024
👀
👨‍💻 THREAD on how data center-related ballot measures are faring tonight data centers are on the ballot in ~40 localities across 13 states, including proposed bans, votes on specific projects, and recalls of pro-data center politicians
122
👀
One of the most concerning developments in open-weight AI has been the rise of vendors of "abilterated" models offering cyber capabilities with no guardrails @joinFAI's FLD is seeking prospective plaintiffs who may have been injured by these systems to support impact litigation
100
v.4 of Who Tests The Frontier AI Labs? What We Know is live. This version adds US CAISSI and Poland’s NASK at OpenAI and Handshake AI at Meta, plus Anthropic’s S-1 disclosure that it indemnifies outside testers and the joint UK AISI–Anthropic investigation of Mythos 5.
1
96
Link for changelog and more → cleverhack.com/who-tests-the…
21
Would love to know why visitors from certain financial institutions have been checking out my Neolab tracker recently.
79
Whoa... I think what is significant is that we're seeing this *outside* of the labs.
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.…
76
If you have any interest in Delaware corporate governance of AI labs, you'll want to take a look...
1
77
The article also talks about Delaware's interest in agent-run companies
11
Oooooof......
it looks like Instinct has hit a wall with Whatsapp, with some users even reporting that their Whatsapp account has gotten banned nearly 50% of the world outside the US & China use Whatsapp...insane distribution moat i'd imagine this would be a no-brainer move for Meta, blocking all other AI agents so they can monopolize global (non-US, non-China) text agents
2
208
Just idly thinking back to early summer on here where someone's anonymous sockpuppet account got mad at me in a thread and said that OpenAI was going to win because they had compute.
be Greg Brockman go to White House dinner to beg your enemies for compute > "h-hey Elon... got any spare compute bro?" Elon calls you out in front of the entire room: "DUDE, you stole my charity! You can't be a huge dick then ask for favors, hello?" Sergey Brin: “He was asking me for compute too. How many grown men do you know that keep a diary? If I did I’d burn it.” LMAOOOOOOOO
5
880
Never in all my days would I have ever imagined seeing the Pope doing the Villanova "V" hand gesture. This is delightful.
Pope Leo just welcomed the Villanova and Notre Dame basketball teams for a special private audience at the Vatican ahead of tomorrow’s Eternal City Tip-Off in Rome.
117
Good lord, who thought this was a good idea?
haha, so apparently OpenAI are about to start training on your Daybreak Blue sessions, with no way to opt out. Literally stealing people's private work, as they do it, is cartoonishly evil. It also will presumably kill Daybreak usage by security researchers. Or at least, it should. The thing OAI have had to swear they are not doing to math researchers, they are now explicitly about to start doing to security researchers. Yet another reminder of the importance of open models.
3
16
837
👀
Some observations on the market for compute: Last night I went to the SF Compute "there's no more compute" townhall talk, Evan and Ethan gave a bunch of great context on the state of the market from where they were sitting and I learned a lot. Interestingly, the first person I talked with as soon as the main presentation wrapped was a guy trying to sell me inexpensive access to capacity on two different B300 clusters. As you can imagine it was an odd experience to be pitched on cheap, readily available compute at the "there's no more compute" function. Separately, I spoke with a few people about TSMC/Intel/Samsung near future production capacity. Apparently it is not common knowledge among purchasers of significant compute that these fabs are not sold out, and that there are significant exit rights available to people who have reserved wafers. N2 is tight, but Nvidia and Apple can make that happen entirely on their own. Shuttle runs are not filling up. If one is properly AGI-pilled one would probably assume that these fabs would be sold out on all nodes to fully pre-paying customers into the early 2030s, yet they are not. I believe Evan and Ethan when they say that the market isn't clearing, it clearly isn't. Exactly why the market isn't clearing is unclear to me. I think there are a bunch of risk dynamics going on right now that are not at all settled, and a bunch of people are at risk of blowing up if compute prices move significantly either up or down. There isn't enough information for me to make any concrete predictions right now, other than that shit is gonna get weird for a bit. Stay safe out there and maybe raise a bigger round sooner than you had planned.
97