
Objective aligned swarms @tenfoldxyz | Prev: AI @medialab @columbia ML @eightsleep
đ˝
Joined July 2011
- Tweets9.5K
- Following995
- Followers3.5K
- Likes35.6K
Pinned Tweet
Round 1: Whoever figures out the single player coordination interface wins (chatgpt)
Round 2: Whoever figures out the swarm coordination interface wins (??)
The latest models are rumored to be âloopedâ transformers, and safety advocates have been afraid theyâd lack interpretability for safety monitoring.
Looks like @GoodfireAI figured out some ways to interpret them that Iâm excited to dig into this weekend.
We got to play with @thinkymachines' new looped model. As it turns out, from an interp pov there's no reason to be intimidated by loops, the model is still full of interpretable structure and mechanisms. Go checkout our new blog post!
1
4
278
JB Rubinovitz retweeted
Today weâre releasing Inkling 2 Mini, a looped transformer, with open weights.
A big reason we're doing this is to let the AI safety community study looping, especially how loops affect chain-of-thought monitorability and how to interpret the computation inside loops. đ§ľ
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/inkâŚ
I don't think this is a consensus, not everywhere.
Humans should fundamentally stay in charge of their destiny, even if we increasingly hand off to AI.
Vibe coding would be a lot less popular if people could still afford to own and renovate their own homes.
almost all of my life now runs entirely on artifacts generated by ai. iâll share more & more examples here as time goes on but here is a fun one.
every morning i wake up to a beautiful little weather report. if i continue to scroll it preps me for anything that might impact my day.
itâs very detailed & has a beautiful presentation layer that i continue to tune.
The last job in the world will be jury duty
Congress is slow by design, as the founding fathers thought slower legislating would invite deliberation and avoid bad laws.
But with frontier AI, this is not enough. Congress needs to move quicker.
@KevinTFrazier proposes a citizens' assembly for AI policy: 500 randomly selected Americans would be selected to draft AI policy with experts as needed.
If the President approves a proposal, it becomes law unless Congress votes it down within a month.
1
4
216
Hospitals are a bigger villain in healthcare than insurers.
They create healthcare monopolies outside of what youâd typically consider âhospitalâ care by buying up practices in every specialty and making everyone sign noncompetes.
Once they have no competitors, they can charge whatever they want.
Hospital prices have more than quadrupled since 2000ârising far faster than other health care costs and everyday expenses like housing and food.
More in our new analysis: on.kff.org/4zYyrvQ

ALT The KFF graph is showing cumulative percent change in the Consumer Price Index for urban consumers from January 2000 to October 2026. Hospital services increased by 314%, overall medical care by 133%, housing by 118%, food/beverage by 109%, prescription drugs by 94%, and overall inflation by 100%.
10
424
TLDR agents are autistic and canât figure out social norms
Can agents figure out the social norms of a new communityâwithout being told what they are?
đ Humans are normatively competent. We watch what others do and what gets corrected, and we adjust our behaviors accordingly.
Introducing NormGym â a benchmark for evaluating normative competence through social interaction.
đ¤ Main takeaway: Agents can sometimes learn from their own norm violations, but struggle to learn from the behavior of others around them. This points to the challenge and the need to build agents that can think and act as members of a community, rather than just pursuing their own goals.
đ Paper: arxiv.org/pdf/2611.08613
đ Project: normgym.github.io/normgym-paâŚ
1
4
418
The influence of AI on labor should be considered a two-player game: technology changes incentives, but workers are clever actors who have many moves at their disposal
I'm in @AmerCompass today talking about one less discussed possibility here: the return of the medieval guilds
JB Rubinovitz retweeted
Californiaâs new AI safety report is the most serious policy proposal for recursive self-improvement from any government Iâve seen. Very exciting!
It recommends the procedural machinery to actually slow down if & when we need to: verified compute disclosures for automated AI R&D, standby authority to cap that compute, emergency shutdown powers, mandatory loss-of-control reporting, and broadened whistleblower protections.
You canât pace what you canât see. If implemented, this would let the state see, and then act.
I hope the legislature moves on this when itâs back next monthâAI developers seem on track to recursive self-improvement within 6-12 months, so we need these mechanisms practiced and working ASAP.
The recommendations from Governor Newsom's recently convened expert panel - Professor Alondra Nelson (Princeton), Jason Goldman, Professor Robert Reich (Stanford), and Professor Gillian Hadfield (Johns Hopkins) - were just released and are a clarion call for urgent action to address risks to public safety from advanced AI systems.
Some of the most notable recommendations:
⢠Emergency shutdown authority
⢠Minimum safety standards for AI developers
⢠Mandatory third party oversight
⢠Authority to cap the amount of compute used for RSI
The next weeks and months will determine whether these recommendations are words on paper, or whether the Governor and legislature can work together to accomplish something historic to protect CA and lead the country on guardrails to prevent catastrophic risks from runaway AI development.
Even people in the business of fine-tuning will tell you there are very few use cases where it makes sense right now.
lol. Nobody who actually deploys agents in prod at scale believes fine tuning tiny models for workflows is beneficial for cost or quality.
The exception is inference commodity workflows (I.e jev type stuff) where youâre running more deterministic chains to categorize emails or something
1
135
Wouldâve been cool if the ChatGPT Astra launch showcased AI natives pushing the model to do extraordinary things instead of zoomers using AI to scam their boomer bosses.
Replying to @tszzl
I will say this guy sandbags a bit unfortunately doesnât show as much initiative as some prior models. on the other hand 5.6 sol was like a Rottweiler just latched onto the RLVR shape and doesnât give up
1
133
Google used to have "20% time" for engineers to work on whatever they wanted. Gmail, Google News, and AdSense all came out of 20% time.
20% time seems to be back in a new way. Great engineers I talk to all have background projects they prompt while they are waiting for their agents to get back to them.
My current background project is trying to figure out why the building down the street from me collapsed and killed 8 people last week. Do you have a background project?
3
331
JB Rubinovitz retweeted
Replying to @doomslide @teortaxesTex
They want me to buy the story that scaling RL from here leads to all the scary replicator like minds. But what if it's all just compiler feedback + IP siphoning? It's all about compute and power and capital at this point. No one has any time for artisanal mind designs.
JB Rubinovitz retweeted
Saw someone say, âThe internet was once a library and they turned it into a store,â and if youâre young I donât think you can imagine how different the web used to be.
JB Rubinovitz retweeted
I fear the rest of the world is going to learn a lot about the dangers of centralization over the next 2-3 years.
JB Rubinovitz retweeted
In the coming machine intelligence biosphere, companies like Cloudflare and Stripe provide the fundamental primitives--the oxygen, carbon, sunlight, etc.--that allow swarm intelligence to flourish. And, importantly, these primitives become more and more useful as intelligence becomes more and more diverse.
Claude Meetings is what unlocks the training data for Claude joining and speaking in conversations with your entire team.
So expect to see that within ~6 months.
Introducing Claude Meetings: add Claude to your teamâs video call and it builds docs, whiteboards, and other artifacts with you while youâre still talking.
Starting to roll out in beta on Team and Enterprise plans.
1
9
407
I think the average personâs fear about âhuman disempowermentâ is actually a fear of further âelite empowerment and average person disempowermentâ
This is a reasonable fear since so far whoever controls the compute and weight controls the models.
it's a good question for those who are worried about "human disempowerment".
What power do you think you have? What power are you worried about losing?
Is it concrete power over your own life that you're afraid of AI taking away... or some kind of control over AI development?
Is it your own power, or the power of some human elite that you imagine to be running things and don't want replaced with AI?
7
330


































