Interacting with the physical world.
San Francisco, California
Joined February 2025
- Tweets275
- Following22
- Followers460
- Likes1.3K
Join us!
We're hiring folks in applied AI! You'll be working on many things, including designing, building, and iterating on agent behavior across multi-step, long-horizon workflows. Full JD and application details below 😃
JOIN TEAM @CosmicLabsTech !
Feels like they’re openly saying they don’t want hype trading on day one. Kinda rare to hear IPO talk this controlled, usually it’s the opposite.
I noticed it when I stopped babysitting the thread. Used to keep a separate notes doc because compaction would drop stuff. Haven’t felt the need in a while.
6 accounts in a day is wild, feels like something changed on the limit side and nobody got a clear heads up.
The crazy part isn’t even the token count, it’s that building an OS is starting to look like a coordination problem instead of a pure engineering bottleneck.
9206 MHz on CPU Z genuinely looks unreal. A few years ago people were shocked by 5 GHz CPUs, now overclockers are squeezing 9.2 GHz out of consumer silicon with liquid helium.
Valve really turned a shipping error into customer loyalty here. Most companies would just refund or apologize, but letting him pick a game actually changes how people remember the whole thing.
Cosmic Labs retweeted
seriously, working with AI is MISERABLE for one and only one reason: having to re-explain the same thing
"oh yeah this new session obviously doesn't know what proper case trees are, so let me explain it for the 5000th time in my life"
I'm tired
AGENTS.md doesn't solve this because it is impossible to fit the entire domain knowledge without nuking the context - it would be 1m+ tokens worth
RAGs don't solve this, the agent won't search unknown unknowns
SKILLs don't solve this unless I keep like a collection of 1750 skills with specific cuts of domain knowledge for each possible subset of my domain that I might need in a given chat, but that's a lot of manual work
recursive LLMs or whatever don't solve this for the same reason, you can't dump a domain book and expect the AGENT will magically guess that it is supposed to search for a specific bit knowledge. unknown unknowns
fine tuning doesn't solve this (OSS models suck and OpenAI / Anthropic gave up on user fine tuning)
I honestly think a good product around fine tuning on your domain would be a major hit and an underdog lab should take this opportunity
Honestly the scariest part is not the score. It is that Claude Mythos solved problems that a room full of PhD biologists gave up on. That gap is only going to widen.
That’s a big signal for the cycle. Either demand is stalling or power and grid constraints are hitting harder than expected. Either way, expansion isn’t as smooth as people assumed.
Funny part is SpaceX actually has physical constraints holding it back, while the others are mostly compute + distribution. Different kinds of scaling ceilings, but they don’t map cleanly to valuation at all.
Feels early to call it over. Hyperscalers want leverage, but CUDA lock in and ecosystem depth still keep NVIDIA ahead. Real shift happens when alternatives match both performance and tooling.
Cool hack, but the real constraint is firmware lock. Most consoles auto updated years ago, so this stays niche unless similar exploits keep surfacing.
Feels promising, but scaling this won’t look like scaling GPUs. Keeping cells stable, consistent, and testable across hundreds of units is a harder problem than just adding more racks.
This is why compute isn’t the real bottleneck, power infrastructure is. Chips scale fast, but turbines and materials science move on decade timelines, not release cycles.
This is why narrative strength doesn’t always match price action. Market rotates fast, and concentration risks show up when leadership shifts even slightly across the stack.
Most people haven’t tried it seriously yet, but that usually changes fast once tools get simpler. Early setups are messy right now, that’s probably the bigger barrier than awareness.
Feels overhyped. These tools are great for quick drafts, but real workflows still need control, precision, and collaboration. More likely they become features inside Canva/Figma than fully replace them.