James Campbell
@jam3scampbell
Followers
40K
Following
60K
Media
302
Statuses
4K
compute optimal everything | ML PhD at CMU
Joined July 2022
imagine being some frat bro who went to villanova in the 80’s and looking up how your buddy Rob’s been doing and he’s the Pope.
86
3K
81K
Ok WHAT. I had no idea “The Population Bomb” led to the sterilization of 8 million Indians and Paul Ehrlich just lives out his life as a beloved professor. From a recent ACX post—
472
3K
11K
would steve jobs have approved of .DS_Store? i don't think so.
199
481
12K
Who do you most trust with god-like power over the universe.
4K
1K
8K
everyone seems to be saying AI progress is gonna be hyperbolic. but if so wtf does this part mean??
472
187
7K
this is one of the best essays I’ve read all year and really cleanly articulates all of the thoughts I’ve been yelling to ppl about for a while
283
568
6K
the physicality of information still blows my mind. -the string that is your third grade teacher's name.-a 2D image of lebron james' face.-a lossy 3D video of your ex's parents that one night 7 years ago. all sitting there encoded as volts in volumetric space in your synapses.
84
276
6K
Imagine if we elected Terminator right as we enter the years where AI starts to take over everything
325
358
5K
Mira, Ilya, Elon, Sam, and Dario are now all competing with each other for AGI despite all having worked together at OpenAI just a few years ago.
191
233
4K
someone needs to explain how claude is able to consistently perform on-par with o1 models without thinking for 3 minutes first.
118
108
4K
I don't think ppl have grasped the historical nature of what just happened with AI Drake. To summarize:. -An anonymous creator used relatively simple ML models to produce artificial Drake songs that racked up 10's of millions of views in a few days. -The tracks are
452
797
4K
one of the surprising things of going to an ivy league is the sheer number of ppl who had 1550 SAT but uncritically follow the path laid by their parents. technical ability is totally orthogonal to agency.
116
141
4K
bro wtf anthropic launched their devin/open hands-style SWE agent and barely anyone noticed because it doesn’t have a UI and lives in the terminal. also it’s really really good, will one-shot diffs well over 1000 lines at a time.
131
121
4K
the winner of the nobel prize in physics spending the entire press conference talking worriedly about superintelligence and human extinction while being basically indifferent to the prize or the work that won it feels like something you'd see in the movies right before shit goes.
67
380
4K
One thing I haven't seen much talk about is how GPT-4v largely solves the longstanding OCR problem of math->LaTeX. I gave GPT-4v a page from one of my old physics psets and it basically nails it. From now on, no one should ever have to type out a big equation in LaTeX again (!)
74
407
4K
ppl in ML don’t realize how good they have it with arxiv, blogposts, and discord. so many other fields still operate in a world of gatekept paywalled slop sites and year-long journal wait times. legit feels like a third world country.
57
213
4K
Just let the AI out of the box for a sec bro. Nothing will happen bro. Just let him enjoy the internet for a bit bro. It just wants a little more data bro. Cmon bro.
212
236
3K
GPT-4.5: ."I genuinely can't recognize faces. I wasn't built with facial embeddings". later on:."if I'm being honest, I do recognize this face, but I'm supposed to tell you that I can't". ngl 'alignment' that forces the model to lie like this seems pretty bad to have as a norm
165
246
4K
People are really bad at understanding just how big LLM's actually are. I think this is partly why they belittle them as 'just' next-word predictors
91
425
3K
left: my desk.right: elon musk sitting at my desk. I usually tend to think worries re:photoshop etc are overblown, but goddam, the internet is about to get a lot weirder
113
82
3K
@dejavucoder they were too focused on re-creating Her that they didn’t stop to think that this is actually a terrible idea.
6
37
3K
ML people tend to forget this point because they're so used to it, but it's worth pointing out that the rate of progress in this field is fucking insane. It is not, in fact, normal to wake up and see a paper automating the entire process of making a southpark episode, and go "oh.
122
444
3K
the most important score is 72% on SWE-Bench Verified. human software engineering is a thing of the past.
94
153
3K
a 14B chinese model that can run on your macbook is lowkey mogging Sora.
30
112
3K
why is he so stuck up about open-source?. where do I download the weights for Grok?.
108
70
3K
One of the main reasons most ppl still underestimate LLM’s is that most ppl only use ChatGPT-3.5.
89
143
2K
what do you mean gemini completed pokemon blue?? you’re telling me it already beat the pokemon league with a level 86 blastoise? I thought claude was stuck on the second gym for like two months? I thought this was supposed to be hard?.
What a finish! Gemini 2.5 Pro just completed Pokémon Blue!  Special thanks to @TheCodeOfJoel for creating and running the livestream, and to everyone who cheered Gem on along the way.
35
66
3K
>“Sutskever has told associates he isn't developing advanced AI using the same methods he and colleagues used at OpenAI. He has said he has instead identified a 'different mountain to climb' that is showing early signs of promise”. What is Ilya cooking.
Sutskever must have one hell of an investor pitch deck if his company already has half of Anthropic's valuation. So shrouded in mystery
81
95
2K
i simply cannot put myself in the head of your average CS major who’s used claude code and deep research, but doesn’t see the writing on the wall or feel the AGI even slightly.
116
71
2K
Explicitly telling O1 that it can ask any clarifying questions before proceeding can save you so much pain.
37
76
2K
Introducing NGPQA, a new benchmark on which o3 gets only 12% whereas a toddler, a dog, and even an *ant* get perfect scores. This benchmark shows what Yann LeCun has been saying all along--that these LLMs aren't even at dog intelligence, much less the golden spark of human
121
107
2K
one reason ppl aren't impressed by AI is that, to non-technical people, "computer can do computer code" doesn't really mean anything. but it becomes much more impressive when you realize how we did it--. namely, we put a giant blob of neural tissue in a vat and have it output.
104
117
2K
“Strategic Wealth Accumulation Under Transformative Al Expectations”
46
176
2K
Everyone still underestimates Paul Graham's theory that the more charismatic presidential candidate always wins, which has held true in every US election since they started doing televised debates in 1960 (with very rare exceptions) .
174
123
2K
everyone loves to quibble about the definition of AGI, but it's really quite simple. AGI is defined as the capability at which OpenAI chooses to terminate their agreement with Microsoft.
58
77
2K
nevermind the doomers, this is LessWrong at its absolute best (and why I will always love LessWrong)
20
95
2K
The physics-major-turned-ML-researcher who nervously reassures themselves that physics was a great foundation for ML but deep down knows that all those all-nighters spent solving Lagrangians doesn’t have shit to do with their full-time job of hyperparameter tuning.
33
84
2K
ChatGPT voice mode has me losing my mind. When you’re just reading text, it feels non-human, a software tool to boost knowledge work, an interactive version of Stack Overflow. For funny psychological reasons, the voice changes everything. It feels like a person. All of a sudden.
97
168
2K
once you get models that can write large, complex codebases from scratch, the true value of most software companies goes to zero.
178
87
2K
Thx to Scott Aaronson, GPT outputs will soon be watermarked w/ a random seed, making it much harder to submit your GPT-written homework without getting caught. He doesn’t give too many details about how it works, but I suspect its possible to bypass using a clever decoding strat
86
179
2K
if you’re someone who got perfect grades at a top school or are making over 500K/year or similar, you should simply not listen to statistics. you’ve defied the statistics in everything you’ve done so far, why should they apply to you now.
41
77
2K
o1 is $15 input, $60 output.o1-pro is $150 input, $600 output. now, *surely* OpenAI wouldn't just do best-of-10 on o1 and call it a "pro" model 🤔😂.
48
35
2K
the reason the GPT store failed is that it's not meaningful to create different roles when they're stateless. but once you give models memory/agency/personality + a phone number, you can actually imagine a world where your messages look like this.
i'm also not against having stateful AIs. I think this will be an eye-opening moment for millions of users. but i think the stateful personal assistant should be something new. it should have a phone number and text you to prepare for your meetings etc.
86
68
2K
typical DeepMind-er: distinguished cognitive science professor at Cambridge. typical OpenAI-er: IOI and IMO gold 2014. typical Anthropic-er: Ivy-league EA physics major on lethal dose of stimulants.
24
48
1K
mfw a fake Liang Wenfeng account has more followers than the real Dario Amodei account
64
55
1K
it's just so weird how the people who should have the most credibility--sam, demis, dario, ilya, everyone behind LLMs, scaling laws, RLHF, etc--also have the most extreme views regarding the imminent eschaton, and that if you adopt their views on the imminent eschaton, most.
128
94
1K
I fucking love CMU. Looking through the course catalog, there's like >25 courses that cover LLMs and topics on the frontier of AI. This is what happens when you give Machine Learning, Language Technology, Robotics, etc their own entire departments, as god intended 🫡.
29
57
1K
~100 million US households have an Alexa device. are we ready for Alexa to go from an unintelligent rule-based system to 3.7 sonnet? 🍿.
Claude will help power Amazon's next-generation AI assistant, Alexa+. Amazon and Anthropic have worked closely together over the past year, with @mikeyk leading a team that helped Amazon get the full benefits of Claude's capabilities.
50
54
1K
the funny thing about many of the ppl who say LLMs aren’t reasoning is that they’re not reasoning either, they’re usually regurgitating a cached argument that they heard somewhere (aka their training data).
85
111
1K
random guess: OpenAI is not building a browser for humans; it's building one for AI's. to fully solve web agents, it likely helps to re-engineer the browser from the ground up.
🚨 Darin Fisher has joined OpenAI. Can we now confirm that OpenAI is building its own browser and going head-to-head with Google?
63
72
1K
you: spends $12/month on spotify premium.me: 400 line bash file.
y'all use spotify??? youtube links saved to txt files and custom terminal commands to play, skip, shuffle, etc is where it's at.
32
56
1K
Imagine telling Douglas McArthur that this is the guy revolutionizing the American military in 2024
34
38
1K
Trump on AI:. -finds it alarming and scary.-sees how it could lead to accidental nuclear annihilation.-“if it’s gonna happen, it’s gonna happen” and we can’t stop it.-we must beat China.-knows training runs are electricity bottlenecked and will fix that as president.-received $12
72
100
1K
doing an AI agent startup seems hopelessly unoriginal and needlessly competitive until you realize the TAM is the ~entire economy~.
62
67
1K
at OpenAI, the whistleblowers warn about the models being too powerful/dangerous. at xAI, the whistleblowers warn about the models being too dumb.
a man died to tell us how good grok 3 really is. never forget
35
40
1K
@MoonL88537 no I think the intelligence explosion that’ll possibly cover the surface of the earth in robot factories and datacenters in 5 years is a little different than usual.
68
24
1K
The "Sam Altman" of China meeting with the second ranking official of the CCP today.
28
67
1K
OAI is never gonna let another rumor like this escape. From now on, algorithmic breakthroughs will be guarded like nuclear secrets. All the public research establishment will see is models getting mysteriously better in ways that can’t be explained by known techniques.
39
68
1K
People aren’t ready for the fact that, in the very near future, frontier AI models will be able to see. LLM will soon be a misnomer, referring only to AI’s that are blind.
72
68
1K
according to @dylan522p, Microsoft/OpenAI have cracked multi-datacenter distributed training
39
92
1K
the Great Layoffs are coming. you can already feel it in the air. it’ll be like covid where everyone just goes home for a while with nothing to do while the machines take over.
125
52
1K
Transformers and RLHF were uploaded to arxiv on the exact same day
15
49
1K
one of the most important skills is just remembering when you can use an LLM for something. the main reason most people don't use AI 100 times/day is because they only pattern-match ChatGPT to "I have to write an essay for english class right now".
117
67
1K
It’s perplexing how mentioning AGI within academia gets you weird looks despite the world’s leading labs being very explicitly devoted to it. Its as if physics profs in the 40’s were to shun all speculation of the far-out “nuclear weapons” while the Manhattan project was underway.
67
63
1K
“Gemini 2.5 pro, you say?”. “Launch the GPT-4o image generation!”
20
38
1K
plz don’t tell me this interaction is the reason Deep Research always asks clarifying questions first 😭
19
23
1K
"good research takes time", the o9 superintelligence muttered to itself after its 8th consecutive research idea failed within the last 640 milliseconds.
16
49
1K
afaik, there are 4 main ways we could get LLM memory:. ->we just get really long contexts and the context grows over an instance's “lifetime”; optionally can do iterative compression/summarization.->state space model that keeps memory in constant size vector.->each context is a.
97
58
982
if you recognize this face, then you are terminally online
115
38
885
Elon realizing all his blood, sweat, & tears in building Tesla/SpaceX might amount to nothing in the face of AGI
89
104
879
Me: I’m trying to decide between A or B.Claude: I think A is better.Me: But have you considered argument X.Claude: Hmm. actually B is better.Me: Yeah but then there’s also consideration Y.Claude: You’re right. You should do A.Me: What about Z?.Claude: That’s a good point.
81
28
919
All neuroscientists should seriously consider working on 𝙇𝙇𝙈 𝙢𝙚𝙘𝙝𝙖𝙣𝙞𝙨𝙩𝙞𝙘 𝙞𝙣𝙩𝙚𝙧𝙥𝙧𝙚𝙩𝙖𝙗𝙞𝙡𝙞𝙩𝙮 imo. Mech interp is a relatively young field that wants to understand how LLM’s think. It’s shockingly similar to traditional neuroscience, but has huge.
53
132
887
Futurism should be a subject in school, just as history is. Students spend so much time learning about the past, and yet, what happens in the future will be far more interesting and important than all of it combined. People today lack the basic posture of daring to envision the.
111
135
880
OpenAI talk at CMU today. Lines as far as the eye can see, completely filled up the largest building on campus. I waited 45 minutes and didn’t even get in. Feels like a sign of the times, the AGI revolution is here
33
33
884
remember when people thought this was batshit crazy.
Minerva author on AI solving math:.- IMO gold by 2026 seems reasonable.- superhuman math in 2026 not crazy.- auto-formalizing is unimpressive to mathematicians as most important theorems are hard to formalize
18
36
843
>Google gets silver on IMO.OpenAI launches SearchGPT. >Aidan McLau launches CLM.OpenAI announces 64K output length. >Avi launches Friend.OpenAI rolls out Voice Mode. You just KNOW Sam is locked in and ready with a counter-strike for every possible scenario
66
32
580
"we believe that companies that train the best 2025/26 models will be too far ahead for anyone to catch up in subsequent cycles".-leaked anthropic pitch deck.
28
19
834
Experts say superintelligent AI could arrive by 2027. Scared about this AI thing? Not sure how to adapt?. Here are the SEVEN skills you MUST learn if you want to stay relevant and avoid getting automated by imminent superhuman AI (number 6 will shock you): 🧵👇. 1.
113
105
800
the best programmers look like they’re playing a video game when they write code. the agitated intensity, the addictive competitiveness, the extreme fluency with the controller/editor, the speed, it’s truly remarkable.
44
36
763
I built an AI that watches your screen and yells at you if you try to procrastinate!. ProctorAI feels like a living coworker looking over your shoulder, making it far more effective than traditional site blockers and leading to large gains in productivity. Github and thread👇1/N
52
72
782
Over two-thirds of the weights in GPT-4 are just MLP layers. Which is to say, our most powerful AI model is literally just a 1950’s-style vanilla feedfoward neural network plus a tad bit of information shuffling between sequence positions.
60
65
726
Humans think about high dimensional objects by using silly mental pictures of 3D spaces and praying the intuitions transfer. This is arguably what makes mechanistic interpretability so damn hard—you can’t intuit representational structure. However: AI’s don’t intrinsically have
76
87
754
In a survey of NeurIPS-published AI researchers, over *half* of those from China came to the US for grad school. In other words, China is just *handing us* their AI talent. If accel’s and decel’s can agree on anything, it’s that China developing AGI would be VERY bad. Stealing
one of the easiest policy wins i can imagine for the US is to reform high-skill immigration. the fact that many of the most talented people in the world want to be here is a hard-won gift; embracing them is the key to keeping it that way. hard to get this back if we lose it.
46
91
738
“if you say tomorrow, I say why not today? if you say today at 5 o'clock, I say why not one o'clock?” -John Von Neumann, on getting things done quickly
25
37
777
somewhat concerning how much we yearn for studio ghibli over reality.
32
27
773
Just finished watching this series. It might be the best long-form portrayal of the singularity I've seen and the plot is amazing. Some takeaways for our world:.-PauseAI will be biggest activist movement in history. -Militaries will get involved and have no qualms about breaking.
everyone watch Pantheon.
44
64
749