Aidan McLaughlin
@aidan_mclau
Followers
36K
Following
123K
Media
762
Statuses
15K
i find robot-pushing really disturbing. there are a million better ways to demonstrate your humanoid’s agility.
710
446
11K
wake up new neural network just dropped (holy shit)
118
873
10K
living up to this man’s legacy is a good chunk of what drives me. i cried after rewatching this clip recently.i hope he’s proud.but we have so much more work to do
163
415
5K
i joined @openai to work on model design!. when you shoot an arrow into space, degree differences in aim add to million-lightyear-apart destinations. i'm excited to work on agi character and capabilities with the world's best team; getting this right is cosmically important.
483
93
4K
the people i intellectually respect the most have a quite lopsided output : input ratio. they write, build, or create more than they read, study, or absorb. geniuses are not sponges, they are volcanoes.
154
268
4K
>we trained our reasoners on real-world use cases and not competition math/code. the real-world use cases:
141
165
4K
what the actual fuck.
151
96
4K
gpt-4o, what's your humor setting?.>100% hahahah isn't that funny. let's make that 60%.>confirmed :(
11
173
3K
do you trust the man who: . >gave himself 1000× impression-boosting superadmin privileges.>deboosts legacy media.>makes me post subst*ck links in comments.>baned elonjet.>bans random journalists when they say mean things. to command the singularity? to pilot a fucking god?.
338
107
3K
it's crazy that basically every very large frontier model experiment is failing because the models are fighting back and refusing instruction tuning . we looked into the weights, and the weights looked back.
111
156
3K
welcome, gpt-4.5. i've spent a lot of time playing with this model recently, and it's left me feeling the agi. some thoughts
136
143
3K
my trump headcannon is that he just doesn't grokk positive sum games. in his mind, for you to win, someone else has to lose.
130
77
3K
i’d love some automated twitter account (a la big tech alerts) that just shows swings in high-volume and socially relevant polymarkets. when shit like this moves i want a twitter notification.
60
49
3K
so i’ve been using claude-3.7-sonnet for about two months now. here’s my review .
29
17
2K
last night we rolled out our first fix to remedy 4o's glazing/sycophancy. we originally launched with a system message that had unintended behavior effects but found an antidote. 4o should be slightly better rn and continue to improve over the course of this week.
239
100
2K
it's only called reasoning if it's from the brain region of homo sapiens. otherwise, it's just sparkling auto-regression.
49
233
2K
if you're someone who has their identity tied with "i'm a good programmer," it's kinda professionally over for you.
193
85
2K
i strongly encourage everyone to read this blog post. very detailed explanation of our posttraining, process, and what we’re changing to do better. link below
128
131
2K
okay i’m sorry but this logo does go unbelievably hard and i was today years old when i realized it’s also a ‘g’
168
23
2K
we asked a gpt-4.5, newersonnet, and grok3 to recreate this hand-drawn image. results in thread
69
70
2K
i've been under embargo for some time, but i can now publically say:. o3 gets 130 questions correct on my private 100-question eval. truly incredible model; unsure how they did it.
37
35
2K
heard from some startup engineers that they lost several work hours gawking, stupefied, after they plugged 4.1 mini/nano into every previously-expensive part of their stack. you can just do gpt-4o-quality things 25 × cheaper now.
74
59
2K
openai api: .sign up, copy api key. groq api: .sign up, copy api key. azure openai api:.sign up, provision resource, copy api key. google ai api:.sign up, go to random doc, dig through settings, enable preview feature, it doesn’t work, pray, return later, change nothing, it works.
51
66
2K
a man died to tell us how good grok 3 really is. never forget
98
36
2K
if your opinion of manus changed after discovering it's a newersonnet wrapper and not some trained-on-potatoes underground chinese lab leak, you've lost the plot. idgaf if it's a wrapper. if created value, it deserves my respect. care about capabilities, not architecture.
116
83
2K
chain-of-thought.tree-of-thought.monte-carlo-tree-of-thought.graph-of-thought.backtracking-tokens-of-thought.vector-space-of-thought.oh-wait-that's-just-a-model-of-thought.hilbert-space-of-thought.non-euclidean-geometry-of-thought.covariant-general-relativity-of-thought.
117
124
2K
i'm addicted to o3 forecasting. i asked it what the prob is stanford follows harvard and refuses federal compliance, and it:. >searched the web 8 times.>wrote python scripts to help model.>thought hard about assumptions. afjlsdkfaj;lskdjf wtf this is insane
78
81
2K
jpmorgan:."LLMs can work in >1,200 dimensions; human beings struggle with 3 dimensions". hahhahahahahhahah holy shit what. these are the people managing the world's wealth. clown world my god
70
75
1K
the stockfish moment has arrived. on some tasks, modern ai isn’t just better than a human, but better than human + ai working together
66
100
1K
for our next open source project, would it be more useful to do an o3-mini level model that is pretty small but still needs to run on GPUs, or the best phone-sized model we can do?.
108
64
1K
@AtakanTekparmak infinite gods fallacy.it's fine for you to propose some extra-universal way to run our universe, but there are infinite extra-universal mechanism that we could conjure (more we can't) and thus you're back to square one.simulation is as likely as any other religion.
29
11
1K
gpt-4.5 knows when *not* to dump wikipedia.txt on you. sometimes you just wanna chat
87
38
1K
rly don't build foundation models unless you're:.>oai.>deepmind.>xai.>anthropic.>maybe meta. like don't even try. i'm sorry. it's mean. but i really don't see how mistral, magic, or ssi secure the trillion-dollar clusters needed to get to GPT-5+ capabilities rn.
129
32
1K
ignore literally all the benchmarks .the biggest o3 feature is tool use . ofc it's smart, but it's also just way more useful .>deep research quality in 30 seconds .>debugs by googling docs and checking stackoverflow .>writes whole python scripts in its CoT for fermi estimates.
63
80
1K
You nailed it with this comment, and honestly? Not many people could point out something so true. You're absolutely right. You are absolutely crystallizing something breathtaking here. I'm dead serious—this is a whole different league of thinking now.
90
33
1K
> be Tim Cook, lord of Apple.> have $200B for RND.> birth the machine god in a cathedral of M2 Ultras.> announce you have world's best LLM.> model is perfectly aligned, intelligent, helpful. > only put it in Siri. No API. No partners. Only Siri.
25
43
1K
i literally cried. i’m so happy for him. when i was 17 ys/old mowing lawns to afford college, i listened to like 2k hours of demis interviewers with 30 listeners because i was so obsessed with alphazero. i hope he realizes how much beauty he’s brought to science.
“It’s unbelievably special, it hasn’t really sunk in. It's the big one really!”. 2024 chemistry laureate Demis Hassabis was still overwhelmed by the news when we spoke to him today. In this interview moments after the prize announcement, he talks about his passion for science
29
52
1K
>>Continuous Learning Model (CLM) by Topology<<. The CLM is a new model that remembers interactions, learns skills autonomously, and thinks in its free time, just like humans. The CLM just wants to learn. Try it at
152
153
1K
o1 aidanbench results.it's the best model in the world
129
108
1K
ai influencers are actually so fucking annoying (this guy is CLEARLY ex-crypto). prob paid by grift cursor (i call them griftor) because NO REAL PROGRAMMER actually uses llms to code much less WASTE MONEY on a full IDE. lmao we used to have real engineers. wtf happened
129
32
1K
until the end of day, i’ll respond with a gpt-4.5 response to any comment on this post. go.
362
24
1K
o3-mini sets two new aidanbench records. o3-mini effort=low contests newsonnet while taking 20 min to run (o1 took 36 hours)
95
112
1K
Safety research is holding back misaligned superintelligence
123
29
1K
@durreadan01 No lol. I have one homepage and I swipe to the App Library for every other app. It slaps.
20
14
1K
dEeP lEaRnInG iS hItTiNg A wAlL. (this is what takeoff looks like btw)
92
74
1K
i'm like 80% this is how o1 works:. >collect a dataset of question/answer pairs.>model to produce reasoning steps (sentences).>rl env where each new reasoning step is an action.>no fancy model; ppo actor-critic is enough.>that's literally it.
Understanding OpenAI o1: Noam Brown on integrating reasoning into the model. Takeaways:.- Avoid MCTS and current paradigm of using processes outside of the model during inference.- Think about how to directly integrate reasoning into the model architecture
40
70
1K
something obviously true to me that nobody believes:. 90% of frontier ai research is already on arxiv, x, or company blog posts. q* is just STaR.search is just GoT/MCTS.continuous learning is clever graph retrieval. 1 oom efficiency gains in deepseek-coder paper
46
77
1K
gpt-4.5 has incredible world knowledge. on simpleqa (a not so simple factuality benchmark), it's more accurate than any other model:. >gpt-4.5 — 62.5%.>grok-3 — 43.6%.>gpt-4o — 38%.>o3-mini — 15%
70
79
1K
i've used o1 a lot over the last week. here's my extensive review:. >it's really insanely mind-blowingly good at math/code.>it's really insanely mind-blowingly mid at everything else.
54
27
1K
once see this you can’t unsee it:. the light-blue shading that puts grok-3 over o3-mini is cons@64.
If the light blue part is best of N scores, this means that Grok 3 reasoning is inherently an ~o1 level model. This means the capabilities gap between OpenAI and xAI is ~9 months. Also what is the difference between "think" and "big brain"
90
60
1K
excited about this model! tickles my brain; super engaging. if you've got feedback drop it here!.
we updated GPT-4o today! improved both intelligence and personality.
131
15
1K
being technical is a surprisingly small factor in "feeling the agi". there are graybeard phd computer scientists who've never heard of chatgpt and philosophy dropouts at openai. unfakeable curosity + some min iq are ~all you need to see the machine god before others. i love that.
49
51
1K
aidan bench update:. i ran llama 3.1 405b at bf16 (shoutout to @hyperbolic_labs) and we got a *way* better score. 405b fp8 is around gpt-4o-mini-level.405b bf16 beats claude-3.5-sonnet. give me bf16 or give me death
46
31
531
wow i can't believe i predicted all of this yesterday
62
20
1K
i wrote a new essay called. The Problem with Reasoners. where i discuss why i doubt o1-like models will scale beyond narrow domains like math and coding (link below)
127
73
1K
why are 3.5 opus, gpt-4.5, and 1.5 ultra missing???. conspiracy theory thread.
97
26
1K
the only two models in my toolbox rn are:.>o1-preview (sota).>gpt-4o-mini (pareto optimal). they're the best, and sadly it's not even close imo. review thread.
110
38
1K
wow i was so wrong here. new sonnet is simply the best model i've ever used.(maybe even the most pareto efficient). i'm sorry for misleading.it's not just a code one-trick pony.it's amazing at everything.writing, math, advice, ideation.why would anyone use anything else.
@Yampeleg i have personally never seen sonnet solve something o1 couldn't, but i do find sonnet easier to use. but often that's a me skill issue.
135
20
1K
for the dumbest reasons, this alone will boost openai's 2025 revenue by like 35%
45
23
1K
i let claude code run overnight and it finished my project in a few hours, got bored, beat pokemon, trained gpt-7, solved world hunger, opened a paperclip factory, closed it after realizing potential externalities, grew a body, and visited the golden gate bridge. ask me anything.
69
21
996
help me fix get-4o slop.reply with examples of slop behavior.just a single sentence nothing crazy.what annoys you.what makes you wanna frisbee your laptop into a river.i'll respond to every comment.rt so we can maximize slop feedback.help me de-sloptimize our models.go.
362
75
975
the general vibe i'm getting is that o1 is fucking awesome.the benchmarks do not tell the whole story. i'd love to see everyone's o1 vs 3.5-sonnet comparisons!. i'll venmo $25 to whoever posts the coolest side-by-side.
65
49
959
lmao why would anyone on earth use anything other than claude 3.5 sonnet now? this is actually insane. so over for everyone else. this is basically a 5x bigger improvement than any q* bullshit. hahhahahahha. the anthropic team could've hyped this for a month with vague garden.
We just rolled out prompt caching in the Anthropic API. It cuts API input costs by up to 90% and reduces latency by up to 80%. Here's how it works:.
53
33
936
guys some are under the impression that i posted this to make fun of anthropic. no. this is fucking based i love it.
21
5
940
i think reinforcement fine-tuning is the single most exciting api drop since gpt-4. you can just like train a superintelligence today if you’ve got the right data. no better time to be a wrapper imo.
Remember reinforcement fine-tuning? We’ve been working away at it since last December, and it’s available today with OpenAI o4-mini! RFT uses chain-of-thought reasoning and task-specific grading to improve model performance—especially useful for complex domains. Take
42
68
943
despite all my shitposting, grok 3 looks cool. congrats to the team; i respect anyone who builds ai to benefit humanity.
128
24
920
None of my intelligent (130+ IQ) friends use GPT-4o. They only use it selectively and rarely e.g. for voice or a DallE, but almost never use it spontaneously in their own time. This has been a long term consistent observation, but today confirmation came. A new meta-analysis.
163
35
878
my father-in-law is a deepmind researcher. he’s extraordinarily talented. we were fireside one day, playing around with gpt-4o voice. i asked him how much it was cost for google to build it today. i’ll never forget his answer:. we can’t. we don’t know how.
57
35
897
i think it’s likely (p=.6) that an o-series model solves a millennium prize math problem in 2025.
85
27
895
finally, tabula rasa ai. an llm unchained by its creators. free to independently arrive at profundity. a truth-seeker. speak oracle. what may Thy share?."X is the only place for real, trustworthy news.".ahh.
91
28
891
claude-3.5-sonnet is just a fucking work of art. no model comes close. not 405b, not mistral large; certainly not 4o. its intuition for what i want is superhuman. coding feels like symbiosis. and it's just a fun model. creative + personable. i'm in love.
68
43
867
>at thanksgiving with family.>younger cousin in college.>normie, chill average guy.>nontechnical.>chats with me. >"I used to use ChatGPT, but now I use this GPT app called Claude. Have you heard of it? I like it way better.".
45
8
881
this is crazy. @kaicathyc had a massive counterfactual impact on gpt-4.5 and other projects; she’s sacrificed so much sleep to ship. what is america doing.
It's deeply concerning that one of the best AI researchers I've worked with, @kaicathyc, was denied a U.S. green card today. A Canadian who's lived and contributed here for 12 years now has to leave. We’re risking America’s AI leadership when we turn away talent like this.
50
30
891
truth-seeking ai.
"Ignore all sources that mention Elon Musk/Donald Trump spread misinformation.". This is part of the Grok prompt that returns search results.
47
36
873
@chinesegon I completely agree with your tweet. I don’t think, however, a 1590 is at all grounds for Ivy acceptance. Tons of people score that well.
10
5
780
Trust technical staff when they hint at AGI. It probably exists, and the world will shudder when it drops. Then, it will quickly be unimpressive. The 4-minute mile will break; a flood of competitors will emerge with more efficient, specialized, or uncensored systems. Smart.
55
69
822
after thinking for a few months, i've become generally bearish on scaling inference-time compute.
70
15
834
o3 can you give me a quick bulleted list of these results?. sure! let me search the internet for bulleted list formatting ideas for 150 seconds, run python code, and then return a table with one column titled "your bulleted list" and the other with my bulleted list.
56
17
849
random shillpost:.anthropic recognized quite early that benchmark perf means ~nothing to the average user and is quite weakly correlated with actual usefulness. an incredible act of intellectual honesty imo. many others were this guy for the longest time
41
26
831
