
@weeklytreemani
iAccount based inUnited States
About this account
- Account based in
- United States
- Connected via
- United States App Store
Account-level information from X, not a live location or the device used for a specific post.
Be Sobeck, not Ted Faro | Eng at some banana cloud
Northeast Regional
Joined October 2022
- Tweets6.7K
- Following1.1K
- Followers1K
- Likes65.4K
Pinned Tweet
if the tokens you generated didn't elevate you out of where you are, of what use were the tokens?
what are small models afraid of?
haiku 5.5 (and haiku 4.5, and luna, all would dilly dally before either descending, or before venturing into a corridor with unknown length...
somehow missed this, cool result...
Hot damn!
Congrats on the AGI, @OpenAI! 😉
kenforthewin.github.io/blog/…
Luxun's alt retweeted
I hope we keep Opus 4.6 around for a long time
my impression of a guy who is in a design meeting and suddenly remembers what the "instrumental" in "instrumental convergence" means:
"is this instrumental to your goal?"
"i think you lost track of the eventual goal with instrumental details"
"let's focus on sufficiently specifying the problem specs because instrumental paths will converge"
if i have a nickle every time a db client whimsy their way out of a performant migration i'd have two nickles
If you can't experiment on something kinda-like frontier LLMs, it's hard to learn things about LLMs, including things about alignment -- that's something Anthropic has explicitly stated.
If you can't *reproduce an experiment* then it's pretty hard to do actual science.
should be a scalable direction.
next we can push back the wings,
next we can merge the wings with the stabilizers,
next we can rearrange the engines for a more concentrated thrust
next we can...

























