ml enjoyer @reductoai | prev math enjoyer @uoft
Joined May 2025
- Tweets 1,016
- Following 569
- Followers 109
- Likes 5,721
2 Photos and videos
it’s interesting to see talk of “solving intelligence” as if this goal has some kind of terminal state.
1
3
108
depending on how you formalize it, various notions of non-computability of such systems start cropping up
1
42
Jun 13
this aged poorly
Jun 12
thinking of ditching codex for claude max 20x just for fable-5
is it worth it guys?
1
2
525
Jun 13
I’m actually at loss for words
Jun 13
I just fell to my knees in a Sweetgreen
176
Jun 13
it’s happening
Jun 13
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.
The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Access to all other Claude models is not affected.
We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.
Read our full statement: anthropic.com/news/fable-myt…
2
129
Jun 13
this is quite an insightful post, the agent trajectories are way more valuable than we realize
4
303
Jun 12
vik’s been cooking!
5
189
Jun 12
it’s going to be interesting for us to look back at this moment in human history.
2
4
279
Karan Brar retweeted
Jun 10
Claude Sisyphus is too on point
Jun 10
the year is 2028. claude infers whether you’ve ever even thought about gradient descent and silently routes your queries to Claude Sisyphus, a model RL’d to maximize engagement while avoiding task completion. you spend your entire UBI token allotment on it without ever realizing.
1
2
10
819
Karan Brar retweeted
Jun 9
They're trying to create and lock in a permanent feudal society, with an elite few having access to power.
31
53
1,091
58,012
Jun 5
RL at scale is a systems problem after all.
Jun 5
very true. infra engineers will have generational run for years and beyond categories.
- model serving / inference infra
- gpu / distributed systems infra
- cloud kubernetes orchestration
- data infra
- eval observability infra
- agent / runtime infra
i still think as ai advances there will be demand for areas which govern bottlenecks underneath like latency, throughput / cost / reliability / data quality and security.
1
27
6,144
Karan Brar retweeted
Jun 4
A real problem with feeling the acceleration viscerally is that current models are really good and it is hard to feel the vibe difference on most individual tasks with new models, even as AIs continue to increase in ability by large amounts (which they actually are doing).
27
12
254
17,796
Jun 5
this reward design is pretty interesting, ai societies solving problems
Jun 4
Imagine a population of machine agents. Each might be strong on certain tasks but fundamentally limited: partial tools, partial observations, finite context, bounded compute.
How can these agents self-orchestrate and self-evolve into stronger collective intelligence to solve tasks beyond any single agent's capability?
Instead of designing the multi-agent system itself, we propose designing the incentives that govern it.
We put agents in an economy. They compete, trade, get wealthy, go bankrupt, and mutate, forming an alive society where coordination and adaptation automatically emerge in a decentralized manner.
1
324
Jun 4
RSI is coming. It will require us to rethink how research is done, and the whole infrastructure around it.
Jun 4
Our internal data shows Claude is accelerating AI development—a possible path to recursive self-improvement, or AI autonomously building a more capable successor.
It’s happening faster than we thought, and the implications deserve greater attention. anthropic.com/institute/recu…
3
354
Karan Brar retweeted
Guys, I beg you to understand that the long term AI frontier is about statistics / information theory and physics. It is NOT about linear algebra.
Dense matrices are the best approach we have rn because of the TPU/GPU hardware and because the math is robust and general purpose.
However, it is inefficient.
And as we optimize towards more faithful intelligence representations sparse networks will dominate the intelligence / energy frontier.
The most important thing is not to be a super physicist information theorist, because only I can be so awesome after all, but to be able to think generally in these terms from first principles. You need to be able to think CONCEPTUALLY in statistics. You need to understand that these matrices are just encoding the necessary information to sample probability distributions.
The 21st century will be the century of information and statistics. Please develop intuition around these ideas.
38
52
710
47,244