andrew gao
@itsandrewgao
Followers
40K
Following
9K
Media
2K
Statuses
10K
@cognition @stanford; prev @sequoia @LangChainAI @pika_labs @nomic_ai; Z Fellow 🇺🇸; views my own
Menlo Park
Joined October 2020
the frontier labs should find new ways to present their eval results. accuracy % really undersells the improvements now that we're closer to 100 than 50. if we assume eval questions are normally distributed in difficulty, that means that as you get higher scores, it becomes
11
4
153
chatgpt branching is soooo useful especially when i've been marinating/curating a gpt5 pro chat for hours
1
0
23
"now let me create a document explaining the key fixes" "now let me create a document explaining how to set up the code" STOP
4
0
7
i cmd F'd my browser history and it turns out that i use chatgpt slightly more than claude. i thought it would be the opposite
2
0
4
anything but personal accountability
Americans want AI companies to be held liable for a wide variety of potential harms. And they're right!
3
1
33
im shaking i thought ashlee vance was a woman for the last eight years
0
0
11
random thoughts
the frontier labs should find new ways to present their eval results. accuracy % really undersells the improvements now that we're closer to 100 than 50. if we assume eval questions are normally distributed in difficulty, that means that as you get higher scores, it becomes
0
1
41
this is definitely too exaggerated but here's what the data looks like with log 10.
0
0
9
time to let GPT-5 pro loose!
saw @josephjojoe_'s scroll friction tweet and had a fun idea: extension that tracks how "far" you doomscroll each day three-shotted with Sonnet 4.5 in @windsurf!
1
0
5
saw @josephjojoe_'s scroll friction tweet and had a fun idea: extension that tracks how "far" you doomscroll each day three-shotted with Sonnet 4.5 in @windsurf!
0
0
16
difficult historical fact that no matter how many people are alive on earth at a given point there’s been the same small number, maybe 1500 or so, that are actually changing the world in any substantial way
45
16
313
the kicker is that jp morgan seems to be suing to not have to pay the $115M in legal fees, and i assume that ms javice is going to respond to that with more legal fees
when jpmorgan acquired her company, there was apparently a clause in the contract that said they would cover her legal bills so when they sued her, she used that clause to spend $115,000,000 of their own money to defend against them Jpmorgan lost 175+115=290M 🤣
3
0
25
Twitter has been so much more pleasant since the bluesky-type people left
The CEO of Bluesky posted an anodyne joke about bringing down the temperature on her website, and the replies are just incredible. The second you let these people in, they eat you alive.
1
0
17
when jpmorgan acquired her company, there was apparently a clause in the contract that said they would cover her legal bills so when they sued her, she used that clause to spend $115,000,000 of their own money to defend against them Jpmorgan lost 175+115=290M 🤣
283
1K
17K
lol instead of saying they used "gpt-4o" they phrased it as "generative artificial intelligence (AI) large language model (Azure OpenAI GPT – 4o) based tool chain"
Let me summarize: one of the world's largest consulting companies uses genAI to write reports and then can't even select a good model that makes significantly fewer mistakes (they used 4o instead of o3 that time)?! "In the updated version of the report, Deloitte added reference
1
0
19
my conspiracy theory is openai put out the claw machine to collect teleop data
3
0
72
how does @ycombinator decide how many startups to fund per batch? (is it just based on quality [if 1000 amazing startups apply, 1000 can get in] or is there a quota/limit) caveat: fall '25 data is incomplete
0
0
3