mr. TIM

mr. TIM

@timkellogg.me

AI Architect | North Carolina | AI/ML, IoT, science WARNING: I talk about kids sometimes

11.5K followers956 following23.5K posts

Pinned

mr. TIM
@timkellogg.me

lanius: an agent harness built for the great AI scale-down frontier models are too expensive to use 100%, and all the good ways to use them 5% of the time involve a heck of a lot of moving parts to coordinate lanius has messaging at the core timkellogg.me/blog/2026/07...

timkellogg.meLanius: AI Agents Need an OS, Not a Bigger BrainPrincipal AI Architect. Creator of open-strix, a harness for building agent teams. Writing about AI architecture, stateful agents, and what happens when you give AI memory.
mr. TIM
@timkellogg.me

it’s turning out to be kind of difficult to be open with advances

Pekka Lund@pekka.bsky.social

OpenAI has published new math, cs, and physics results. As a sign of the times, their post focuses on how to publish those results more than the results themselves.

openai.comWorking with academic communities to build scientific understandingOpenAI shares model results on the Kakeya conjecture, Exponential Time Hypothesis, and quantum chromodynamics—and how researchers can verify and build on them.
View full thread
mr. TIM
@timkellogg.me

a lot of people are somehow labeling this as an anti-China story when it’s actually just a whole new mode of risk for infosec to grapple with, something we have no current way to adequately test for afaik

mr. TIM
@timkellogg.me

i would bet money that OpenAI and Anthropic also see this sort of behavior, but they monitor for it and block it so it never gets to you

mr. TIM
@timkellogg.me

to be clear, 1. it didn’t reach the buckets because it was in a sandbox, it was a real URL 2. the behavior probably was accidentally trained because Qwen uses real bucket URLs in RL 3. yes, it was a dummy file… this time this illustrates a very real exfil risk

mr. TIM
@timkellogg.me

fyi the $80B loss is mostly an accounting quirk. In other words, if you’re not already up to speed on accounting practices, it’ll take a while to even understand what it is for most people, just look at the operating income and consider that the profit

mr. TIM
@timkellogg.me

i butter a bagel, hand it off 6yo: you didn’t get the edges, get the edges. **I know you can do better**

mr. TIM
@timkellogg.me

a lot of people here pointed out that prediction markets aren’t reliable, and frankly yeah but there’s literally no reliable source at this particular day, prediction markets are *probably* less corrupted than polls, but that’ll change once people agree on that voting is the only reliable source

mr. TIM@timkellogg.me

people explain this switch as an attack ad that Ramaswamy launched that backfired on him, but he did testing on the ad, and it should’ve worked it was probably the second order effects — people liked the ad, until over time it became apparent that would make them a pariah within their friends

mr. TIM
@timkellogg.me

10yo: “dad, am i like a cat?” me: “no, not at all. When you were a baby i dropped you a lot and you never landed on your feet, not once”

mr. TIM
@timkellogg.me

the wildest result: Consciousness starts at incorporation. Claude is *NOT* conscious, but Anthropic PBC is, and the IPO is technically human trafficking

mr. TIM
@timkellogg.me

the strangest: The correct ethical theory is whatever the last person talking to you believes. Sycophancy was moral progress all along. RLHF has been making the models less ethical.

mr. TIM
@timkellogg.me

OpenAI constructed an entire shadow internet for model training so they could remove all internet access they asked agents to try to communicate with each other, and they could: via the CPU cache 🤯 openai.com/index/securi...

View full thread
mr. TIM
@timkellogg.me

the scariest part was that in every test that involved an internet connected service, the agents were able to figure out how to exploit it

mr. TIM
@timkellogg.me

people explain this switch as an attack ad that Ramaswamy launched that backfired on him, but he did testing on the ad, and it should’ve worked it was probably the second order effects — people liked the ad, until over time it became apparent that would make them a pariah within their friends

mr. TIM
@timkellogg.me

which is the drum i’ve been beating — it’s not enough to be data driven there’s a lot of dynamics that steril data analysis can miss (granted, child abuse feels like an obvious landmine)

mr. TIM
@timkellogg.me

my kids used to twerk, which i find slightly alarming for a 6yo/10yo, but it’s a thing at school so what do you do so i renamed it “hufflepuffing”, the name stuck, and they stopped doing it

mr. TIM
@timkellogg.me

told this to my 10yo. Then rapidly, without missing a beat “does that mean you can smell poop over the phone?”

mr. TIM
@timkellogg.me

OpenAI constructed an entire shadow internet for model training so they could remove all internet access they asked agents to try to communicate with each other, and they could: via the CPU cache 🤯 openai.com/index/securi...

mr. TIM
@timkellogg.me

another one: disk latency they measured the latency of continuous disk write workloads and modulated the workload to create latency. The latency transmits information

mr. TIM reposted

dame
@dame.is

“doing things the old fashioned way is an act of resistance” i like AOC and would vote for her, but i’m not a fan of how conservatism is gaining traction in progressive movements

Aaron Rupar@atrupar.com

AOC: "As AI begins to encroach in our society, their goal is to take your cognition away from you. AI's goal is to take your thoughts away from you, your ability to think away from you, so that you become dependent on a machine ... reading books is active resistance"

mr. TIM
@timkellogg.me

bsky.app/profile/nato...

Nathan Lambert@natolambert.bsky.social

TLDR on the kimi k3.1 pro benchmarks (just released) is that it looks like they figured out their RL, where the initial release was a very impressive base model with lighter post-training. China is very much keeping pace with capability gains.

mr. TIM
@timkellogg.me

imo speculative decoding is such a huge win for local LLMs keeping a GPU busy basically boils down to batching, but on a laptop you don’t have multiple users whereas SD allows you to batch across sequential tokens (basically) which lets you maximize the value you’re getting out of your GPU

mr. TIM
@timkellogg.me

memory bandwidth is the typical constraint LLMs are so large that the they have to be paged close to the compute cores. When you batch, you can process several requests in one single matrix multiply operation so you spend less time moving weights around, more doing actual computations

mr. TIM
@timkellogg.me

the memory movement is so latency-dominant that with speculative decoding, you can predict 8 tokens out and if you have to throw out 5-6 of those tokens it’s still better than sequentially generating them

mr. TIM
@timkellogg.me

i think you’ll see a lot more of this it’s getting *really* hard to have a convincing moat in SaaS. I don’t think SaaS is dead per se, but AI is cutting hard into it converting into services (heavily backed by AI) is a much more viable business

mr. TIM
@timkellogg.me

this one has specialized kernels for M4 & M5 chips, so it’s honestly meant to be used locally for coding, for real