chronically early
Joined August 2021
- Tweets 1,065
- Following 174
- Followers 159
- Likes 3,188
25 Photos and videos
Jun 13
praying for this to get ratiod
We fully support @POTUS and @SecWar in prioritizing national security and the security of our warfighters, DIB partners, critical infrastructure, international partners and allies. Some things are simply more important than revenue cycles, clickbait, and pre-IPO valuation. America First. Always. 🇺🇸
53
Jun 11
like AI for research rn ~=
the infra (AX/design) you use to keep track of state over long runs, compute u throw in at each stage (extreme: cc workflow for reflection), and resources you have access to in agent-native form factor (papers).
objective setting
agent harness
Jun 11
would wager the specific novel thing about the system (bc its the simplest place where there's ludicrous amounts of alpha) is
exposing to arxiv/semantic scholar properly done with a great CLI/python library
and some gepa-like outer loop. maybe with single item minibatch, and a cracked reflective stage that has access to the literature.
viola! gains
(caveat: i'm pulling this outta my ass lmao idk how they do it)
25
Jun 11
would wager the specific novel thing about the system (bc its the simplest place where there's ludicrous amounts of alpha) is
exposing to arxiv/semantic scholar properly done with a great CLI/python library
and some gepa-like outer loop. maybe with single item minibatch, and a cracked reflective stage that has access to the literature.
viola! gains
(caveat: i'm pulling this outta my ass lmao idk how they do it)
1
1
111
Jun 11
i guess you'd also actually wanna combine this with something like flywheel/ARA ?
bc you'd want to fully understand the graph and naive gepa doesn't give u that for free.
1
33
Jun 11
my wellbeing has gone down so significantly lmfao
30
Jun 11
p helpful when designing agent interfaces to have fable ask haiku questions to see if they make sense
1
1
24
Jun 11
gives extremely rapid feedback! in a couple seconds. and fable is good at asking the right questions!
eg: after those qs, instant change based on haiku's misunderstanding:
1
21
Jun 11
85k 2.5k anthropic credits/org 🤨
Jun 11
We’re launching Claude Corps, a national fellowship program matching people early in their careers with US nonprofits.
We'll teach 1,000 people to use Claude, and pay them to use AI to advance their hosts’ missions.
anthropic.com/claude-corps
26
Jun 11
price war brings a big ass smile to my face, we gon be eating this summer boys
Jun 11
No details on how big the price cuts will be beyond 'drastic', but the WSJ says this is being considered by OpenAI in anticipation of similar cuts in token pricing that Anthropic is preparing to announce. So they are headed for a price war by the look of it.
1
73
Jun 9
only startup i've seen tackle the most important bottleneck for shitting out AI software, with two very capable people behind the wheel!
Today we’re launching @modaicdev @a16z @speedrun 006, the fastest way to render your judgement into reliable decision automation.
1
55
Jun 9
terrible to do this without any clear wording on edge cases and what the limits are.
if i'm writing a text optimizer, is that ok ? wb environments ? data labeling in long conversations ? curating data mixes ?
what does "frontier" even mean in this context? where are the lines
When Fable 5 is used for frontier LLM development, it does not notify the user and instead limits the model’s capabilities through methods such as prompt modification, steering vectors, and PEFT.
Anthropic estimated that this would affect approximately 0.03% of traffic.
1
49
Jun 5
aw man u know me always coding up my Tiananmen square simulators hehe
Jun 4
DeepSeek V4 “improved” the code and said nothing happened in Tiananmen Square on June 4, 1989
32
May 28
helluva “speed up your timelines” kinda year. vibe wise, not numbers wise.
19
May 14
overview of the agent issue triage view i made *5 months ago*.
still haven't seen a better UX for debugging agents!
loom.com/share/70386ccee5a34…
47
Apr 13
i think the mythos stuff makes sense cuz isnt superhacker getting out of their box something basically all of the ai safety people were/are worried abt?
58