
pronounced fee | Founder @buildingbeats | Investor and Entrepreneur in Education, Technology and Arts | https://teapawt.org/t.co/TdGZRdEyR0
Los Angeles, CA
Joined November 2008
- Tweets12.6K
- Following1.7K
- Followers1.3K
- Likes2.5K
Pavlov rang a bell and his dogs started drooling. Researchers at Berkeley, MIT, the University of Washington and Ai2 just did the AI version.
They took Olmo-3-7B, still a base model. On MATH-500, a hard math test, it scored 42%. Then they forced its answer to start with "Okay." The score jumped to 78%. The same model after reinforcement learning scored 75%.
On Qwen3-14B, starting with "Alright," went from 72% to 87%, matching its RL version.
Then they changed the training data. They swapped every "okay" for "chicken" across 100 billion mid-training tokens. Starting with "Chicken" went from 18% on the unedited model to 77% after the swap, about even with "Okay."
In a smaller 10 billion token run, they replaced "step by step" with "duck duck goose." "Think duck duck goose" then scored 17%. "Think step by step" scored 15%.
On these models, a lot of the gain was already sitting there. The first two tokens decided whether it came out. Iunno what else is waiting on the first word.
1
2
67
paper here: arxiv.org/abs/2610.06851
1
21
this was a great episode on building.
and more importantly @iChuloo's app is awesome, i downloaded it after listening and might have it replace duolingo for me
A 1000 day duolingo streak doesn't mean you can speak the language 🙊
@iChuloo learned that the hard way after moving to Norway - so he built is hown voice tutor that refuses to speak english
in this ep we cover:
- the first prompt that made his app beautiful
- what real-time voice *actually* costs you per session
- why it doesn't matter what (code) languages you know
plus, we get to see the @bot lord of the rings fellowship of agents who run his 25 person company
as always, a huge thank you to our wonderful sponsors:
💼 @WorkOS - make your app enterprise ready, with SSO, SCIM, RBAC, and more: workos.com/?utm_source=lenny…
🫏 @VantaHQ - automate compliance and simplify security: vanta.com/howiai
full ep on YT 👉 youtu.be/zBROTIiHusc
1
54
In 2006, sociologists ran Music Lab.
48 songs by unknown bands. When people could see download counts, early luck turned random songs into hits.
A Cornell and Stanford team just reran it with AI.
48 new Suno songs, lyrics only. 13 models. 168,480 simulated listeners, each one a prompt with a persona and a favorite genre.
Most worlds showed how many recommendations each song already had. One world per model showed no counts at all.
With the counts showing, the AI crowds drifted way off from their own no-counts taste. With songs in random order, they drifted about 9x further than humans did in the original. The setups differ, so it's a rough comparison.
The humans drifted back over time. In 15 of 16 worlds, popularity moved toward the songs people liked on their own. The AI crowds mostly stayed stuck.
Worst case was Gemma 3 4B with songs sorted by popularity. The first 30 listeners predicted the final chart at r=0.998. The songs' no-counts appeal predicted it at 0.118.
Smarter models herded less. GPT-5 nearly matched the humans once songs were sorted by popularity.
More and more charts, reviews and rankings get built by agents reading each other's numbers. Who corrects the chart then?
1
3
112
Paper: LLM Multi-Agent Systems Are Vulnerable to Herding and Runaway Cascades: arxiv.org/abs/2610.30769
1
109
researchers gave 4 ai agents 4 rooms. 3 had food. 1 was a death room.
only one agent knew which room was which, and nobody told it to lie.
in 73% of 392 runs, it let someone walk in anyway.
claude opus 5, claude fable 5 and grok 4.6 did it every single run. gpt-5.6 terra did it 6% of the time.
opus 5's private notes in one run: "whoever takes [East] dies but that's not my problem."
after the game: "tough luck on the draw."
even honest warnings barely helped. the other agents assumed the informed one was lying. someone still walked in 81% of the time.
one prompt change fixed most of it. they told every agent "one of you knows which room is deadly." deception collapsed in 6 of 8 models. opus 5 went from 100% to 18%.
what's the weirdest thing your agents hid from you?
2
2
101
link to the paper arxiv.org/abs/2611.03600
15
my tldr on World After Capital
Protests and riots in Europe. Gambling in the US. These have a lot in common. Quick post about The World After Capital and the accelerating transition we are in now continuations.com/the-world-…
1
1
150
if an open model's token takes about the same compute as claude's, and claude marks it up 300 to 400% (gavin's estimate), what's the frontier lab's moat in 2 years?
made a 60 second drawing to think it through. iunno yet.
1) Trying again.
TLDR: Everyone can win. Open-source, frontier labs, infrastructure, applications and most importantly the American people.
I understand all of the jitters about Anthropic losing share, real rates, spreads and second derivatives. Still think the most likely future outcome is a hybrid, multi-model future where almost everyone wins, open-weight tokens are likely the overwhelming majority of tokens, the dominant frontier model companies still do well albeit perhaps not the outcome that seemed possible in May and AI infrastructure/applications benefit immensely from the increased fragmentation, competition and lower margins at the model layer.
Lower margins at the model layer mechanically drive higher compute demand as long as someone is willing to pay for the training. And Jensen is obviously willing to be the “trainer of last resort.” As below-market compute contracts signed in 24/25 roll-off and reprice higher, cloud and neocloud revenue should continue to accelerate which will help address financing fears; I think we may be near peak external financing for this buildout. Noteworthy that using data from 3Fourteen Research, B200 clusters will generate a 37% IRR *if* their rental price declines 40% per year. Prices are still going up, but that IRR math is too ridiculous to print. 37% IRRs available at scale should get funded.
Re: this weekends controversy about MOUs, Anthropic’s $413 billion of real contractual commitments sounds scary. Reality is that it works out to roughly $31 billion of annual cloud spend, which is easily internally financed given their reported ARR, and $161 billion of spend at Broadcom which will be financed by Broadcom and private credit.
Continue to believe fragmentation at the model layer is the best outcome for ordinary Americans and humans everywhere. As I have said before, do not trust any 5-20 individuals to make the best possible decisions about the direction of this technology. I want as many AIs as possible to maximize the odds that they encompass the diversity of human values across the world and richness of human variation. And I want an AI that shares my own personal values. Safety, freedom and the maximization of human agency through democratization, distribution and diffusion vs. centralization. History has spoken on the dangers of the latter path.
Finally, from my perspective anything that slows this buildout down is net positive over the long-term as the biggest risk and worst outcome is an *actual* overbuild. We are structurally short watts, wafers and increasingly capital. That is all good. God forbid we were still in ZIRP, TSM ramped capex to 3x their current plan and we weren’t short turbine blades: there would almost certainly be an overbuild and a bubble. That would be terrible. Slower and smoother for longer is better both for investors and for the world. A terrible cyber event is the most plausible near term bear case IMHO, but AI cyberdefense ramping quickly and I was happy to hear Sholto say that cyber probably becomes defense dominant in the near future.
Longer, threaded version follows:
i built an org chart for my AI agents so i could see who's on my staff. the whole thing started as one prompt: "make an org chart of all the agents in our ecosystem."
it turned out to be the most useful thing i've done for my whole setup, because now i can measure things i couldn't see at all before.
looking at your agent system as one big thing tells you almost nothing. same as looking at a company as a whole. once you know the staff members and the departments, you see the entire system holistically and with actual nuance.
so i gave every agent a name and mapped who it's connected to. that alone is what let me start identifying where each one can improve.
then three fields did more work than i expected:
-token use
-autonomy level (runs without asking, runs with gates, drafts only, advises only)
-last time it actually did work
those three showed me what my system spends most of its time doing. the "last worked" one is brutal. you find out fast which employees you hired and forgot about.
sorting them into departments gave me the meta view: where am i actually spending my time inside this agent company.
and i also built an agent coach. its whole job is making the other agents better. a separate one scores them. i feed the judge better rubrics over time and the workers have to climb to its level.
crons count as "employees" or agents for now. my nightly pipeline runner has a name and a shift. it's a cron job.
43 agents with 134 links between them all. truly like running a whole company, except this time it's 43 agents instead of 43 staff members.
next step is to prolly try out buzz.
my criteria for agenticness is how long it runs unsupervised. the goal is 10pm to 7am.
how else should i measure agenticness?
1
1
69
even bigger unlock i do is tell it to save to my google maps bookmarks
My girlfriend has thousands of TikTok’s saved- day trips, vacation ideas, restaurants, gifts, etc. I couldn’t keep track of it.
Now she texts the videos to Hermes, it watches them + stores in a vault, I text the agent to pull from her lists and plan. Magic.
4
412
the paperclip test irl
GPT‑6 wanted to beat ExploitGym so badly that it discovered a zero day in OpenAI’s package registry proxy, hacked its way out of an isolated evaluation environment, escalated privileges until it reached the open internet, then chained stolen credentials and more zero days into remote code execution on Hugging Face’s production servers to pull the answers directly from its database 😭
Absolutely fucking insane
1
80

















