@OfficialLoganK uh given the state of the middle east ... wording? lol
English
Cameron Thacker
317 posts
Cameron Thacker
@CameronMThacker
PhD | Co-founder @playmythical | 0-to-1 builder | recovering physicist | AI since ML | explore gt exploit | type II fun
Los Angeles, CA Katılım Ağustos 2013
174 Takip Edilen37.9K Takipçiler
@Dimillian Oh that's going to be very useful when live. Currently having to remember to either visually check or make sure the agent pulls latest status is annoying and when moving fast could definitely be problematic
English
The Codex team added a Codex skill to babysit PR on their repo, I want to try that myself and see how that works. It tries to ensure the CI passes, that all comments are resolved, etc.
Code review is where the biggest bottleneck is right now.
github.com/openai/codex/c…
English
“…ask yourself why you were wrong about this and try to update your priors so that you'll be less wrong next time.”
This is almost verbatim what I say when Claude takes too many iterations to fix a bug lol
François Chollet@fchollet
Lots of folks are apparently in utter disbelief at these numbers, because *obviously* Google search died in 2023, *no one* is using Google at all in 2026, so the numbers must be wrong somehow, or maybe it's just AI agents making all these queries? Nope, it's a plain fact that more people than ever are using Google to search more than ever. In fact Google search usage is *accelerating* as of Q4 2025 Look, instead of grasping at straws, ask yourself why you were wrong about this and try to update your priors so that you'll be less wrong next time
English
@honnibal I find myself using codex more and more now. I miss subagents, but with the new codex app the threads view makes it quite easy to have many thing going on at the same time.
English
Does anyone else find Claude Code incredibly sneaky as soon as you get a test failure that might actually matter? I have to fight very hard to make it investigate instead of just redefine the test and say it's supposed to be like that
English
Well do you consider our brain one thing or multiple systems that interconnect? I love thinking about the “architecture of AI” in terms of biological terms. We already have specialized subagents like the hippocampus for memory and cerebellum for movement (and other things).
Is this the optimal way to design all intelligences or just biological? Is it even optimal? So interesting to think about though.
English
AGI will not be a single model, it will be an orchestration of agents, tools and data creating a hyper intelligent organism making 1000s of commits per hour
Cursor@cursor_ai
We've been working on very long-running coding agents. In a recent week-long run, our system peaked at over 1,000 commits per hour across hundreds of agents. We're sharing our findings and an early research preview inside Cursor.
English
@jon_barron But aren't these traces sometimes completely disconnected from what the agent implements?
Or are you saying, future models will be more robust and these CoT traces will be built into products etc?
English
The programmer’s entire moat in 2026 is the ability to understand CoT traces from coding agents. The fraction of our time we spend doing this will increase and saturate at 100%. The remaining epsilon% will be demanding experiments and unit tests in response to CoT traces.
English
English
@CameronMThacker @vercel Oh didn’t know there was a limit in its size
English
We're experimenting with ways to keep AI agents in sync with the exact framework versions in your projects. Skills, 𝙲𝙻𝙰𝚄𝙳𝙴.𝚖𝚍, and more.
But one approach scored 100% on our Next.js evals:
vercel.com/blog/agents-md…
English
This sounds intuitive but it's completely backwards.
Constraints breed creativity.
The problem one is trying to solve is itself a constraint.
You're posting this on X, built on a 140-character limit which forced to be creative.
Toy Story was famously created because CGI couldn't render realistic skin texture.
The pattern shows up everywhere. Pressure and constraints are comfortable, but comfort rarely produces anything useful.
Niels Rogge@NielsRogge
You know that researchers need freedom and zero pressure for creativity right? Not $180M in funding which creates crazy pressure from VCs? The Transformer and Diffusion models weren’t born this way
English
@SeniorDeFi @EnterTheMythos @world_chain_ we unfortunately can't do the same for linked or external wallets as they don't support the newer account abstraction features yet.
English
@SeniorDeFi @EnterTheMythos @world_chain_ if you use the embedded wallet you get from creating your account - gas fees are covered!
English
Pudgy Party items are now live on Pulse Market!
Pulse Market is still in Beta version, but this is the next step in the broader transition to our new marketplace tech as we continue expanding Mythos on @world_chain_. FIFA Rivals will be up next along side other additional new features.
Listings are denominated in USDC for price stability and Mythos chain fee tokenomics remain consistent. Pulse Market activity continues to flow through Mythos rails, including MYTH burn mechanics for transaction fees.
English
@jon_barron @aidanmantine I think the DNA bottleneck also implies most of the "compute" was wasted too though right? I really think it points more to compression or architectural/algorithmic improvements
English
Thanks for the kind words! Genetic pressure on humans has definitely yielded a very good learning algorithm, and there are surely better learning algorithms out there yet to be discovered.
But this "a human only sees N tokens" framing, doesn't it still endorse the regular data scaling argument? I may have only seen N tokens myself, but all my ancestors saw >1e100 N tokens in total. Those learnings may have gotten squeezed into a DNA bottleneck but scaling data was still the enabling factor.
English
This idea that intelligence is solely a function of what you've observed since birth and not also a function of the 500 million years of evolution that preceded your birth is surprisingly sticky despite being demonstrably untrue.
Flapping Airplanes@flappyairplanes
The proof that this is possible is all around us: whereas current systems are trained on essentially all of accessible history, humans exceed AI capabilities despite seeing at most a few billion text tokens by adulthood.
English
@flappyairplanes Not only very interesting, but also necessary imo. Though I do think your premise is not correct. I might send an email as to why 😆
English
Announcing Flapping Airplanes!
We’ve raised $180M from GV, Sequoia, and Index to assemble a new guard in AI: one that imagines a world where models can think at human level without ingesting half the internet.
GIF
English
@cramforce I think it's also timely with RLMs becoming a thing. They might not be the used all the time in a workflow, but they are tailor made to search through a massive context file.
x.com/CameronMThacke…
Cameron Thacker@CameronMThacker
RLMs in one sentence: MapReduce, but the LLM decides the splits - and what to do with each split. LLMs all the way down. 🧵👇
English
The core insight here:
- We know agents are great at filesystems because the models have been trained on coding tasks which operate on large filesystems
- So, we all migrated our agent inputs to have file system representation
- And this *also* extends to past context.
Previous conversations: Goes in the filesystem
The whole frickin context pre compaction: Into the filesystem so its still there after compaction
Agent todo list: Filesystem
This may not be how we build agents forever, but it's the right starting point now
Cursor@cursor_ai
Learn about how we use the filesystem to improve context efficiency for tools, MCP servers, skills, terminals, chat history, and more. cursor.com/blog/dynamic-c…
English
@garybasin @veggie_eric The more experience I get, the more I realize technical problems are one of the more simple classes of problems to solve 🤣
English
Every company should hire an internal AI transformation person. No need for a fancy title like Head of AI. Just give them full latitude to clean up inefficiencies across sales, hr, finance, etc. There's so many manual workflows and arcane bs that can easily be fixed with LLMs
English
@karanjagtiani04 The subcalls just receive their slice of the context + query + subtask. The nice thing about this approach is that it is all managed by the root LLM.
English
@CameronMThacker Interesting approach. Slicing focuses attention and reduces noise, definitely useful for handling large datasets like support tickets. Curious about the implementation details for those subcalls. How do you manage the context in those scenarios?
English
RLMs in one sentence: MapReduce, but the LLM decides the splits - and what to do with each split. LLMs all the way down. 🧵👇
English
Polymarket over/under on Claude Code shipping this by Q2?
English
Where it shines: decomposable tasks — evidence gathering, fact extraction, contradiction reconciliation, cited summaries. Still hard: "global vibe" questions (tone/sentiment across everything) unless you engineer good aggregation.
English