Hi everyone yes, I left OpenAI yesterday
twitter.com
twitter.com
This is insane
https://github.com/karpathy/nanoGPT
And it's accompanying video series:
https://karpathy.ai/zero-to-hero.html
Another example (although I honestly don't remember if he made this one between jobs) is: https://github.com/karpathy/micrograd
[1] https://youtube.com/playlist?list=PLkt2uSq6rBVctENoVBg1TpCC7...
https://www.youtube.com/watch?v=NfnWJUyUJYU&list=PLkt2uSq6rB...
Being a popular AI influencer is not necessarily correlated with being a good researcher though. And I would argue there is a strong indication that it is negatively correlated with being a good business leader / founder.
Here's to hoping he chills out and goes back to the sorely needed lost art of explaining complicated things in elegant ways, and doesn't stray too far back into wasting time with all the top sheisters of the valley.
Edit: the more I think about it, the more I realize that it probably screws with a person to have their tweets get b-lined to the front page of hackernews. It makes you a target for offers and opportunities because of your name/influence, but not necessarily because of your underlying "best fit"
if only we compensated that knowledge properly. Youtube seems to come the closest, but Youtube educators also show how much time you have to spend attracting views instead of teaching expertise.
> It makes you a target for offers and opportunities because of your name/influence, but not necessarily because of your underlying "best fit"
That's unfortunately life in a nutshell. The best fits rarely end up getting any given position. May be overqualified, filtered out in the HR steps, or rejected for some ephemeral reason (making them RTO, not accepting their counteroffer, potentially illegal factors behind closed doors, etc).
it's a crappy game so I don't blame anyone for using whatever cards they are dealt.
Actually for all the attention that the top Youtubers get (in terms of revenue), the reality is that it's going to be impossible to replace teaching income with popular Youtube videos alone.
Based on what I've seen, 1 million video views on Youtube gets you something like $5-10K. And that's with a primarily US audience that has the higher CPM / RPM. So your channel(s) would need to get to about 6 million views per year, primarily US driven, in order to get to earning a median US wage.
to get 6m views you need to make one video a week that gets 114k views 6000000/52 = 115,384.61.
Something I've been thinking a lot about is the transition into post scarcity and how we need to dramatically alter the incentive structures and payment allocations.
I've been asking this question for about a decade and still have no good solutions: "What do you do when x% of your workforce is unemployable?" (being that x% of jobs are removed without replacement. Imagine sophisticated and cheap robots. Or if needed, magic)
This is a thought experiment, so your answer can't be "there'll be new jobs." Even if you believe that's what'll happen in real life, it's not in bounds of the thought experiment. It is best to consider multiple values of x because it is likely to change and that would more reflect a post scarcity transition. It is not outside the realms of possibility that in the future you can obtain food, shelter, and medical care for free or at practically no cost. "Too cheap to meter" if you will.
I'll give you two answers that I've gotten that I find interesting. I do not think either are great and they each have issues. 1) jobs programs. Have people do unnecessary jobs simply so they create work wherein we can compensate them. 2) Entertainment. People are, on average, far more interested in watching people play chess against one another than computers, despite the computer being better. So reasons that this ,,might,, not go away.
This can be self-fulfilling.
In an organization beyond a certain size, there will be more almost-adequate-fits than there are leadership positions. This could be about like a recognized baseline which seems like it really needs to be scrutinized closely to see exactly who might be slightly above or below the line.
Or in a small company where there is not any almost-fit whatsoever, imagination can result in an ideal that is equally recognizable, but also might not be fully attainable.
Either way it could be OK but not exactly the best-fit.
If good fortune smiles and the rare more-than-adequate-fit appears anywhere on the horizon though, it's so unfamiliar they fly right over the radar.
Here's a gem of educator. Check out his other videos.
He uses elegant hand-drawn notes rather than Manim - although 3blue1brown's open sourced visualization library is beautiful too, I think this makes it extra impressive.
3b1b's main selling point is the extreme level of polish on his visualizations - something that takes a lot of time (money) to develop
the sad part is that it takes extreme luck to make it on yt. i wish educating skills counted for more but unfortunately they don't, really.
The algorithm is very good at rewarding good content. (It’s also good at rewarding other things, but that is besides the point)
Perhaps some automation/ai combination where you feed it learning videos and it helps create the "other" content.
MOCs are great for access, but they are not, and definitely should not be treated as, replacements. That I am certain will have a net negative result. I'm in grad school and there's something I tell students on the first day:
> The main value in you paying (tuition) and attending is not just to hear me lecture, but to be able to stop, interrupt, and ask questions or visit me in office hours. If you are just interested in lectures I've linked several on our website from high quality as well as several books, blogs, and other resources. Everyone should all use these. But you can't talk to a video or book, but you can to me. You should use all of these resources to maximize your learning. I will not be taking attendance.
I'm sure many of you have had lectures with a hundred students if you went to a large school (I luckily did not). You're probably aware how different that is from a smaller course. It's great for access and certainly is monetarily efficient, but its certainly not the most efficient for educating an individual. MOCs are great because they increase the ability of educators to share notes. We pull from one another all the time (with credit of course), because if someone else teaches in a better way than I do, I should update the way I teach. MOCs are more an extension of books. Youtube is the same, but at the end of the day you can't learn math without doing math. Even Grant states this explicitly.
Strikes a balance between sounding engaging and soothing at the same time.
When my kiddo was a sophomore in HS he decided that he wanted to be an engineer, and I thought that it would be really good for him to learn calc- my feeling was that if he got out of HS without at least getting through Calculus he'd have a really hard time.
So _I_ learned calculus. I started with basic math on Kahn and moved to the end of the Calc AB syllabus. I have, like, 500K points there. And I've watched a whole lot of STEM on YT.
Yesterday I finished a lab with Moritz Klein's Voltage Controlled Oscillators, where I was able to successfully understand the function of all the sections in the circuit.
I've been trying to follow Aaron Lanterman's Georgia Tech lectures on analog electronics.
The issue is that I have other stuff going on in my life. Like, my son studies more than I work at my full time job.
And I don't really have the pressure on me to learn the more advanced math that he's using. In fact, in the couple of years since he graduated HS, I've not really found a use for calc in my day-to-day work on any of the technical things I've done (mostly programming) and so I've lost a lot of it.
So, by contrast, my son who will be graduating as a BS in ME in May, has a far better and deeper understanding of the engineering material than I do.
And it's not just a time issue- I quit my programming job last summer because I have just enough work as a musician to pay the rent, which leaves me plenty of time to do stuff. And it's not that I don't know how to learn at a college level- I taught in an English Dept for 8 years and quit a PhD in the humities ABD.
That's all just my experience.
I love STEM (and trades education) material on Youtube, but I really think that it's missing something to think that you could get " a better undergraduate education in STEM on youtube".
A teacher can usually adapt the content depending on its audience, I would not teach the research in my field at the same level to professionals, PhDs, master students, bachelor students, amateurs, or even school students.
If what I'm teaching is fairly complex, it requires a lot of background that I could teach, but I would not have the time to do so, because it would be to the detriment of other students. So, while I usually teach 'from scratch', depending on my audience I will obfuscate some details (that I can answer separately if a question is asked) and usually I will dramatically change the speed of the lessons depending on the previous background, because I need to assume that the student has the prerequisite background to understand at that speed fairly complex material.
As an example, I gave some explanations to a student from zero to transformers, it took several hours with lots of questions, the same presentation to a teacher not in the field took me 1h30 and to a PhD in a related field took 25 minutes, the content was exactly the same, and it was from scratch, but the background in the audience was fairly different.
To give a relevant example, graph theory concepts can be found both in so many real-world systems but also in programming languages and computer systems.
Just a friendly heads-up, it’s “bee-lined.”
I normally wouldn’t point that out, but “b-lined” could be read to suggest the opposite of your intention; a lower priority, a la “B-list celebrity.”
He should teach more and take it seriously
/s
Plugins were a failure. GPTs are a little better, but I still don't see the product market fit. GPT-4 is still king, but not by that much any more. It's not even clear that they're doing great research, because they don't publish.
GPT-5 has to be incredibly good at this point, and I'm not sure that it will be.
As the popularity has exploded, and ethical questions have become increasingly relevant, it is probably worth taking some time to nail certain aspects down before releasing everything to the public for the sake of being first.
Altman saga, allowing military use and other small things step by step tarnish your reputation and pushes you to the mediocrity or worse.
Microsoft has many great development stories (read Raymond Chen's blog to be awed), but what they did at the end to other competitors and how they behave removed their luster, permanently for some people.
That would actually increase their standing in my eyes.
Not too far from where I live, Russian bombing is destroying homes of people whose language is similar to mine and whose "fault" is that they don't want to submit to rule from Moscow, direct or indirect.
If OpenAI can somehow help stop that, I am all for it.
My experience is limited. I got it to berate me with a jailbreak. I asked it to do so, so the onus is on me to be able to handle the response.
I'm trying to think of unethical things it can do that are not in the realm of "you asked it for that information, just as you would have searched on Google", but I can only think of things like "how to make a bomb", suicide related instructions, etc which I would place in the "sharp knife" category. One has to be able to handle it before using it.
It's been increasingly giving the canned "As an AI language model ..." response for stuff that's not even unethical, just dicey, for example.
If that's the foundation your luster is built on - then it's not really ridiculous.
GPT popularized LLMs to the world with GPT-3, not too long before GPT-4 came out. They made a lot of big, cool changes shortly after GPT-4 - and everyone in their mother announced LLM projects and integrations in that time.
It's been about 9 months now, and not a whole lot has happened in the space.
It's almost as if the law of diminishing returns has kicked in.
Add to that a company environment that seems to be built on money-crazed stock option piling engineers and a CEO that seems to have gotten power-crazed.. I mean they grew far too fast I guess..
Note well: I haven't actually used it myself, so I'm speculating (guessing) rather than saying that this is how it is.
Idk, I just tried Gemini Ultra and it's so much worse than GPT4 that I am actually quite shocked. Trying to ask it any kind of coding question ends up being this frustrating and honestly bizarre waste of time as it hallucinates a whole new language syntax every time and then asks if you want to continue with non-working, in fact non-existing, option A or the equally non-existent option B until you realise that you've spent an hour trying to make it at least output something that is even in the requested language and finally that it is completely useless.
I'm actually pretty astonished at how far Google is behind and that they released such a bunch of worthless junk at all. And have the chutzpah to ask people to pay for it!
Of course I'm looking forward to gpt-5 but even if it's only a minor step up, they're still way ahead.
Initially it felt like the singularity was at hand. You've played with it, got to know it, the computer was taking to you, it was your friend, it was exciting then you got bored with your new friend and it wasn't as great as you remember it.
Dating is often like this. You meet someone, have some amazing intimacy, then you get really get to know someone, you work out it wasn't for you and it's time to move on.
That on top of my own experiences, and heaps of anecdotes over the last year.
> How would they honestly be getting worse?
The models behind GPT-4 (which is rumored to be a mixture model)? Tuning, RLHF (which has long been demonstrated to dumb the model down). The GPT-4, as in the thing that produces responses you get through API? Caching, load-balancing, whatever other tricks they do to keep the costs down and availability up, to cope with the growth of the number of requests.
--
If it's a conversation with "format this loose data into XML" repeated several times and then a "now format it to JSON" I find often it has trouble determing that what you just asked for is the most important; I think the attention model gets confused by all the preceding text.
edit: as pointed out, this was indeed a pretty esoteric example. But the rest of my attempts were hardly better, if they had a response at all.
Here are some hilarious highlights: https://twitter.com/Suhail/status/1757573182138290284
Google DeepMind is still an AI research powerhouse that is producing a ton of innovation both internal and publicly published.
For lots of applications the speed/quality/price trade offs make a lot of sense.
For example if you are doing vanilla question answering over lots of documents then 3.5 or Mixtral are better than GPT4 because the speed is important.
For some advanced reasoning you're 100% right, but many times you're doing document conversion, summarizing, doing RAG, in all these cases GPT 3.5 performs as good if not better than GPT 4 (we can't ignore cost and speed) and it's very hard to distinguish between the two.
- price per thing you use it with matters (a lot)
- making sure that under no circumstances are the involved information leaked (included being trained on) matters a lot in many use cases, while OpenAI does by now have supports that the degree of you being able to enforce it is not enough for some use cases. In some cases this is a hard constraint due to legal regulations.
- geo politics matters, sometimes. Being dependent on a US service is sometimes a no go (using self hosted US software is most times fine, tho). Even if you only operate in the EU.
- it's much easier to domain adapt if the model is source/weight accessible in a reasonable degree, while GPT-4 has a fine tuning API it's much much less powerful a direct consequence of the highly proprietary nature of GPT-4
- a lot of companies are not happy at all if they become highly reliable on a single service which can change at any time in how it acts, the pricing model or it being available in your country at all. So basing your product on a less powerful but in turn replaceable or open source AI can be a good idea, especially if you are based in a country not at best terms with the US.
- do you trust Sam Altman at all? I do not and it seem short sighted to do so. In which case some of the points above become more relevant
- 3.5 level especially in combination with domain adoption can be "good enough" for some use cases
Good.
I have no idea what's really going on inside that company but the way the staff were acting on twitter when Altman got the push was genuinely scary, major red flags, bad vibes, you name it, it reeked of it.
I was surprised and touched by their loyalty, but maybe I missed something you noticed.
They were loyal to money, nothing to be touched by.
Yeah putting people out of work on an industrial scale is probably gonna have a pretty big effect on global GDP
Just imagine what valuation OpenAI would have as a grid monopolist combined with nVidia, ARM, Intel and AMD! Hundreds of trillions of dollars!
Personally, the chat UI is the main limiting factor in my own adoption, because a) it’s not in the tool I’m trying to use, and b) it’s quicker for me to do the work than describe the work I need doing.
Though the TTS side has some trouble switching languages if only single words are embedded. A single German word inside an English sentence can really get butchered. More training needed on multilingual texts (and perhaps preserving italics). But anyways this is really only an issue for early language learning applications in my experience.
In my experience, stopping to talk even for a moment already makes it submit. This makes a real conversation with pauses for thought difficult, because of the need to hurry before it cuts off.
Multi languages would be so useful for me.
I'm not particularly interested in having it outright program for me (other than say to sketch how to do something as inspiration, which I'll rewrite rather than copy) because I think typically I'd want to do it a certain way and it would take far longer to NLP an LLM to write it in whatever syntax than to WhateverSyntaxProgram it myself.
But used as autocomplete, it's definitively a time saver. Most of us read faster than we type.
> Personally, the chat UI is the main limiting factor in my own adoption, because a) it’s not in the tool I’m trying to use, [...]
though I haven't tried it through some combination of it the effort to set it up & it not particularly appealing to me anyway. The best it could possibly be would be like pair programming (back seat) with someone who does things the same way as you, and reviewing their code. I read faster than I type, but probably don't review non-trivial code faster than I type it. (That's not a brag, I just mean I think it's harder and takes longer to reason about something you haven't written, to understand it, and be confident you're not missing anything or haven't (both) failed to consider xyz.)
If ChatGPT doesn't have product-market fit, what actually has?
My guess is it isnt, these systems are hard to trust, and the rhetoric "were aiming for AGI" suggests to me that they know this and AGI might be the only surefire way out.
If you tried to replace all of a devs duties with current LLMs it would be a disaster, making sense of all that info requires focus and background thinking processes simulataneously which i dont believe we have yet.
Overall a chatbot like GPT-4 may be useful, but not that useful as it stands.
If you can write well, it's not really going to improve your writing. Granted, you can automate a few tasks, but it does not give you 10X or even 2X improvement as sometimes advertised.
It might be useful here and there for coding, but it's not reliable.
GPTs are also pretty good, and being able to invoke them in regular chat is also handy, but the lack of monetization and the ability to easily surface them outside of chatgpt is also kind of a problem. These problems are more fixable than the plugin issue IMO since I think the architecture of plugins is a limiting factor.
It is practically unusable and I'll likely cancel paid plan soon.
Convinced hey do it on purpose.
I could understand the sentiment when you think that OpenAI is really doubling down just on LLMs recently, and forgoing a ton of research in other fronts.
They’re rapidly iterating though, and it’s refreshing to see them try a bunch of new things so quickly while every other company is comparatively slow to release anything.
The same nonsense happened with Apple, where like a month after they first released Apple Watch people were yelling "What's next???!!!! Apple is dying without Steve Jobs!"
I don't think it's hugely surprising given the massive hype. No doubt OpenAI are doing impressive things, but it's normal for the market to over value it initially as everyone tries to get onboard, and then for it to fall back to a more sensible level.
Transformers were invented with the support of Google (by the researchers, not by Google).
Open community has been creating better and better models with a group effort; like how ML works itself, it's way easier to try 100,000 ideas on a small scale than it is to try a couple of ideas on a large scale.
I've been anti-Google for a while now so I'm not biased.
I don't think openAI have this sown up.
Relative to his level of fame, his actual level of contribution as far as pushing forward AI, I’m not so sure about.
I deeply appreciate his educational content and I’m glad that it has led to a way for him to gain influence and sustain a career. Hopefully he’s rich enough from that that he can focus 100% on educational stuff!
That blog post inspired Alec Radford at Open AI to do the research that produced the "Unsupervised sentiment neuron": https://openai.com/research/unsupervised-sentiment-neuron
Open AI decided to see what happened if they scaled up that model by leveraging the new Transformer architecture invented at Google, and they created something called GPT: https://cdn.openai.com/research-covers/language-unsupervised...
"In fact, I’d go as far as to say that
The concept of attention is the most interesting recent architectural innovation in neural networks."
Even the HN discussion around this had comments like "this feels my baby learning to speak.." which are the same comparisons people were saying when LLMs hit mainstream in 2022
Take CS 231, for example, which stands as one of Stanford's most popular AI/ML courses. Think about the number of students who have taken this class from around 2015 to 2017 and have since advanced in AI. It's fair to say a good chunk of credit goes back to that course.
Instructors who break it down, showing you how straightforward it can be, guiding you through each step, are invaluable. They play a crucial role in lowering the entry barriers into the field. In the long haul, it's these newcomers, brought into AI by resources like those created by Karpathy, who will drive some of the most significant breakthroughs. For instance, his "Hacker's Guide to Neural Networks," now almost a decade old, provided me with one of the clearest 'aha' moments in understanding back-propagation.
He contributed to pushing forward AI, no “actual” about it. The loss of a great educator should be viewed with just as much sadness as the loss of a great engineer.
He did pioneering research in image captioning - aligning visual and textual semantic spaces - the conceptual foundation of modern image generators. He also did an excellent analysis of RNNs - one of the first and best explanations of what happens under the hood of a language model.
IMO governments, like websites, should be boring but effective, focused on small day to day improvements, not all flash and empty marketing chasing cultural trends...
But try convincing a democracy that politicians should be paid more.
What do you know about his work?
He's been leading the vision team at Tesla, implementing in the field all the papers that were available in the subject of autonomous driving and vision (he explicitly wrote that). He has not published about it surely due to obligations with Tesla.
Are you sure about your perspective?
https://scholar.google.com/citations?view_op=view_citation&h...
So, he has made theoretical contributions to the space, contributions to prominent private organizations in the space, and broadly educated others about the space. What more are you looking for?
Agree he had a decent overall track record at Stanford, but that’s not how tenure works — it might have got his foot in the door as an assistant professor somewhere. He chose a much more lucrative path.
He lead a team of one of the most common uses of DNNs, if that isn't 'pushing AI forward', I think you're confused. It's certainly pushing it forward quite a bit more than the publishing game where 99% of the papers are ignored by the people actually building real applications of AI.
I'd say that that his work on AI has been significant and his ability to teach has contributed to that greatly.
I don’t see that as an issue though, just a natural consequence of his great work in teaching neural networks!
I mean, I don't know why people still try to devalue educating the masses. Anyone who's had to knowledge share know how hard it is to make a concise but approachable explanation for someone who knows relatively little about the field.
In addition, he's still probably in a standing well above the 80% mark in terms of technical prowess. even without influencer fame I'm sure he can get into any studio he wishes.
I'd agree with that, however I've always wondered how easy it is for folks at that level to get hands on keyboards and not wind up spending their days polishing slide decks for talks instead.
Does anyone here know?
You can check his GitHub: https://github.com/karpathy
Anybody have better info than my idle guess?
Television content for children is often called 'Children's Programming'
Isn't that is what Rabbit R1 is? https://www.rabbit.tech/
I really hope he’s next gig is at an actually Open AI company.
And no one seems to have heard what Ilya Sutskever's status is at OpenAI as well.
He might be talented, but if he can’t be trusted, he needs to go.
Ironically that may be exactly what Sutskever thought about Altman.
Building business requires latticework of talented people doing job properly. And building a system of checks and controls for less trusted people.
I'd imagine if one was fully onboard with the AI/LLM commercialization train, there's no better spot than OpenAI right now.
—Carver Mead, 1979 (employee at Xerox PARC), discussing why Xerox needed to focus more on adopting integrated circuits into the computers they had already developed, instead of continuing to just make increasingly-obsolete copiers.
[1] https://en.wikipedia.org/wiki/Mead%E2%80%93Conway_VLSI_chip_...
https://en.wikipedia.org/wiki/Carver_Mead
Learning about the interconnectedness of all this historic intellectual "brain theft," keeps me excited for an AGI-future, post-copyright/IP. What are we going to accomplish [globally] when you can't just own brilliant ideas?!
----
"Xerox's top executives were for the most part salesmen of copy machines. From these leased behemoths the revenue stream was as tangible as the `click` of the meters counting off copies, for which the customer paid Xerox so many cents per page (and from which Xerox paid its salespersons their commissions). Noticing their eyes narrow [at R&D's attempts at asking to market their computer, one] could almost hear them thinking: 'If there is no paper to be copied, where's the `click`?' In other words: 'How will I get paid?' "
—Michael Hiltzik's "Dealers of Lightning" (p272)
In 1979, I doubt copiers were 'increasingly obsolete'; I'd expect the market was growing rapidly. Laser printers, email, the Internet, didn't yet exist; PCs barely existed, and not in offices. Almost everywhere would have used typewriters, I suppose.
>Laser printers, email, the Internet, didn't yet exist
Actually, all three did; the latter was in the form of ARPANET [to be technical, not "The Internet"].
> Actually, all three did; the latter was in the form of ARPANET [to be technical, not "The Internet"].
True, but a technicality. Very few people knew they even existed, and they had zero impact on Xerox copier sales.
You can create and encourage small teams, but then they need to coordinate somehow. Coordination & communication overhead grows exponentially. Then you get all the "no silos" guys and then its all over..
The guy has many tens of millions of dollars most likely.
If OpenAI, Tesla and Google cannot retain him, then probably nobody can. Probably he'll be doing YouTube videos all day long.
To attract someone at Karpathy's level you would need a project that is both wildly challenging (and yet not the typical startup "challenging" because it's a poorly thought out idea) and requires the kind of resources (compute, data, human brains in vats, etc) that would make your place look far more interesting than OpenAI.
But, hardest of all, you would need startup founders that could tame their egos enough to let someone like Karpathy shine. I haven't talked to a Bay area startup founder in a while who wouldn't completely squander that kind of talent by attempting to force their own half-baked ideas on him, and then try to force him out months later when he couldn't ship those poorly thought out products citing lack of "leadership".
It seems from the outside like he locked Tesla into a losing, NN-only direction for autonomous driving.
(I don’t know what he did at openai.)
(Eg Mercedes has achieved level 3 already).
- it costs too much
- it's ugly
- humans have only vision
TESLA Engineers wanted LIDAR badly, but they have been allowed to use it only on one model.
I think that autonomous driving in all conditions is mostly impossible. It will be widely available in very controlled and predictable conditions (highways, small and perfectly mapped cities etc).
And about Mercedes vs Tesla capabilities, it's mostly marketing... If you're interested I'll find an article that talked about that.
But seriously, right now with full attention to LLMs, and many brains, there is no single key person. The question 'who said it first' isn't that important for the progress. With experts leaving OpenAI will gradually loose it's leadership. Others will catch up. Which is good in general, no one should have monopoly on AI. I wish it was that easy with hardware too...