It gives some context on the contributions of each of the authors.
About Shazeer, from the article:
Shazeer’s joining the group was critical. “These theoretical or intuitive mechanisms, like self-attention, always require very careful implementation, often by a small number of experienced ‘magicians,’ to even show any signs of life,” says Uszkoreit. Shazeer began to work his sorcery right away. He decided to write his own version of the transformer team’s code. “I took the basic idea and made the thing up myself,” he says. Occasionally he asked Kaiser questions, but mostly, he says, he “just acted on it for a while and came back and said, ‘Look, it works.’” Using what team members would later describe with words like “magic” and “alchemy” and “bells and whistles,” he had taken the system to a new level.
As a hacker, I kinda like naom's code. I was had to implement a TC MoE kernel, and stumbled upon his code from [tensor2tensor](https://github.com/tensorflow/tensor2tensor/blob/master/tens...) and i think "alchemy" is justified. Dude writes some beautiful kernels.
He also saw LLM would replace search before anyone else, and that is something to look at the Lamda or GPT-1's output and think: yeah this will answer all of our questions one day.
There's no doubt about Noam's abilities. But I read through that code, and struggle to see its 'magic' or 'alchemy'. Can you elaborate what you find especially good about that code? (You may assume GPU kernel programming knowledge on my end.)
To me the magic Noam moment was when he came to my team and said "that cluster has a bad node in it, but this other one doesn't" and we had to spend like a week tracking down a single bad processor out of thousands.
Unrelated to the particular code above. There's a difference between writing code about or adjacent to a proven idea vs writing code in uncharted territory. I suspect that is what happened here. It's the same thing with say music and art. A lot of people today can play Chuck Berry.
It's a good point. Though I do wonder if the magic he casted was more at the conceptual level (intense belief on a set of primitives that ought to work) more than the code itself. Even by 2018's standards, the Tensorflow code above doesn't really look that impressive. It's hard to judge based on those past standards, though. But, wonder if somebody who knows more than me can elaborate.
And, of course, "interesting" means "interesting to dang"; whether apparently technical sources have apparently received PR training is therefore not "interesting". Drop the "my preferences are actually just objective truth" routine for once. Why is it so painful to admit you curate this site by preference?
It's not a question of painful - I'm happy to "admit" what's true, as best I can, and not what's not true. Let's see if we can sort that out a bit in the present case.
HN is certainly curated - I've been "admitting" that since the day I got outed as a mod here:
But we try hard to do the curation by principle, not by personal whim. What principles? Really there's just one: intellectual curiosity—we try to feature what enhances that and dampen what degrades it [1]. From that starting point, though, you can derive lots of other principles. Probably the most important is that snark and indignation are bad for HN (especially in combination!) because they drown out curious conversation. That's all that you need to see why I posted that reply to the GP; no personal preference required.
The current case still seems very heavy on personal preference. Principles application is subjective as we are all human. I found the comment as interesting as the quote it is answering.
These are preference-based but you're pretending they're objective. I find _your_ comments to be full of snark and indignation more than any you respond to, but of course you won't agree. (But because you don't agree, that makes me objectively wrong, I know.)
"Tonal arguments are ways of, frankly, policing working class ways of communication, and covering them in elite preferences." - someone smarter than the average HN commenter.
Mods routinely lie and ignore their own code of conduct, okay favorites, and act highly partisan. This site is far from any sort of meeting ground of ideas.its a narrowly tailored set of mostly statist propaganda.
The "bells and whistles" label sounds more dismissive / perjorative to me. An odd, and not a particularly nice, thing to say. Makes me wonder how the "magic" and "alchemy" terms were intended in this case, also.
signull is more of an anonymous sh*tposter than a known industry insider, but I think this does capture the sama contribution to OpenAI very well. At least from an outsider who follows this stuff based on vibes.
That twitter story isn't anything unique to OpenAI or Google, it's just classic "big public corp vs private startup" culture. Once you have to worry about the SEC, shareholders, antitrust, regulations, lawsuits, etc. it's very, very difficult to avoid turning into "big corp" culture.
Sama, and any other founder, will always have a difficult fight against bureaucracy, and once you let a little bit in, the bureaucracy's sole purpose becomes to grow itself.
Google is facing a legitimate innovators dilemma here. It makes sense to have all this process when youre protecting a $4.5 trillion golden goose. The tragedy here is that one predictable outcome of this situation is google deciding to considerably cut research funding when they figure out it just serves to bootstrap future competitors.
This is when it makes sense to split your business up into multiple smaller businesses. The government should be doing this via anti-trust but they have dropped the ball there so, at this point, the corps really need to just do it to themselves to better compete.
Or maybe just have your R&D teams focused on doing R&D with zero corporate interference. Staff it with personal assistants whose only job is to ensure the researchers have whatever they need and are never bothered with meetings or other corporate shenanigans. The assistants could then be the proxies to management to provide feedback to management, but only on best effort and still staying the fuck out of the way of the researchers.
Google bloat gave us transformers. Apple bloat gave us a usable touchscreen only, pocket computer (famously an entire org within Apple had developed an iPod-based approach that was competing with what was released)
The leaps forward need bloat. A startup can execute on specific vector direction way better.
Now back to your point, what did X deliver with its lean ops? It seems that it needed 2 bailouts (one from xAI, and one from space X)
I disagree. It's not about the culling, it has never been, and actually, it makes things worse. You spend countless hours and tons of money recruiting talented people not to lay them off because you don't want a bureaucratic org.
If the issue is inefficiency, tons of meetings, too much team alignment etc, then that's the issue that you need to tackle, and these issues can already appear in a 50-100 employee company. Sure, that's an easy problem to solve with a smaller size but unless you hired people for no reason, these people have a very specific set of problems to tackle and are often, in these companies, the best in class to tackle them, culling half of the company isn't going to make things better.
It's impossible to disambiguate but advertiser tools, brand safety, targeting, reporting etc all need a lot of ongoing effort. If it gets harder to advertise effectively on Twitter, those dollars can very easily go elsewhere.
Eh, what has X/Twitter delivered since the cull? It’s basically in maintenance mode. Which is fine if that’s what you want to do, but Google and Apple definitely don’t (and I’m glad for that)
Has been in more of a maintenance mode with a multiple of those people. If anything, the pace of the product has improved. Regardless of what you think about Musk, the company he bought was a bloated mess.
Google is competing with nvidia (TPU), AWS (GCP), Netflix (youtube), Tesla (waymo self driving), OpenAI (Gemini), Microsoft (Workspace), Apple (Android)....
I know some pretty wealthy people. They are very aware of those who are 10x wealthier than them. If Noam has 1B, he is probably pretty aware of those that have 10B. He's met them and seen their properties, scope, and powers. Likewise, they are thinking about those that have 100B, and those are thinking about Elon, who now has "four commas."
Even back then Mike Judge said he had to tone down the absurdity he saw on fact-finding trips to Bay Area. He said no one would believe how absolutely stupid so much of all of it he saw was.
The gap between reality and satire was apparently already very small back when the the show was written. The creator, Mike Judge (who also created Beavis & Butthead, and Idiocracy) had worked in Silicon Valley as a developer and based the show on what he saw. Apparently it was very popular with SV insiders precisely because it was so accurate.
Judge also consulted with various teams at places like Google; I worked with one of the guys who provided details that later showed up on the show (as well as many plushies). He didn't watch the show because "it hit too close to home"
Noam Shazeer was one of the lead authors of the seminal paper "Attention Is All You Need", which introduced the transformer architecture. (From Wikipedia)
The architecture was Shazeer's, but the rough idea came from Jakob Uszkoreit who initiated the project.
Uszkoreit wanted to build a more efficient/scalable language/seq2seq model that could take advantage of GPU parallelism (replacing RNNs which were the main approach to sequence modelling at that time).
Uszkoreit's insight was that although language appears sequential, it is in fact really part parallel part hierarchical, as can be seen by linguist's sentence parse trees where at each level there is parallelism/independence between the branches of the tree, with them getting combined at the next level up. This is what gave rise to the idea of a model that consisted of a stack of of parallel processing layers (transformer layers). I believe that attention was also part of the plan from day one, as this had already been proven to be valuable (Bahdanau) with RNN seq2seq modelling.
So, this is what Uszkoreit wanted to build, but by his own account he failed to come up with an implementation that matched or outperformed the prevailing RNN approach that he wanted to replace. At this point, Uszkoreit mentioned the idea to Shazeer, who got on board and eventually arrived at a performant architecture which was then pared back by an ablation process resulting in the initial encoder-decoder Transformer architecture. Shazeer later came up with the mixture-of-experts architecture, and also other optimizations after he left to found character.ai
Curious about others' contributions, such as Vaswani, Parmar, Jones and Gomez, to the paper. What sucks about co-authorship in research papers is that you don't get a clean breakdown of who contributed what to the research paper, and the distribution (in more cases than not) is very much like a pareto distribution.
I'm talking from plenty of group project experience here.
Can you expound on the ablation process? Is that referring to a stripping down of the data or weights or something? Or a stripping down of the transformer architecture structurally? Just curious
You train the model then do a baseline evaluation. Then you evaluate many variants where you have removed or nulled out different layers or chunks of the model. By comparing the performance of those mutated models to the baseline you can learn a lot about the model. What parts don't have much value and can be removed, the location of "functions" or "facts." Etc. Google it.
If you read the Wired article linked elsewhere on this thread, then it explains that. The work was being done by people from the Google Translate team.
It was originally built as a general purpose sequence-to-sequence (seq2seq) model.
The research history leading up to this was interesting - there had been a bunch of work, in various domains, on "autoencoder" architectures used to learn compact representations for things like dimensionality reduction and sequence representation. The idea was to have an encoder-decoder pair, connected by a limited bottleneck representation, with the training goal of the decoder reconstructing the encoder input from the bottleneck representation.
One example of this was to learn a fixed size(!) sequence (e.g. sentence) representation using an LSTM-based autoencoder (LSTM->embedding->LSTM), which at the time seemed rather shocking - the ability to represent a variable length sequence with a fixed size embedding. Equally shocking was that you could use this for machine translation simply by connecting an LSTM encoder for one language to an LSTM decoder for another language.
This type of LSTM->LSTM seq2seq encode-decode architecture for machine translation was then improved by Bahdanau by replacing the fixed size representation with an attention mechanism so the decoder could learn to be more specific about input-output relationships.
This type of LSTM-based seq2seq encode-decode architecture, using attention, is what Uszkoreit et al set out to improve - to make more efficient by using a parallel vs sequential (RNN) architecture. The Transformer was never conceived of as purely for language modelling, or as an "AI" architecture. Later when the usage focused on language modelling (generation, not translation), the encoder was dropped since input and output are the same thing.
Source for this? The notion of attention dates to a content-addressable lookup during sequence alignment (as well as, concurrently, memory lookups in neural Turing machines). Attention had been used in other models, like GRUs and LSTMs with attention. The Vaswani et. al. paper did not introduce attention, just removed everything _but_ attention (and FFW) from the network. Are you claiming the "critical idea" of removing the GRU and LSTM parts and just keeping attention was "truly" Noam's?
At some point in late 2017 the paper was updated with this additional detail:
Equal contribution. Listing order is random. Jakob proposed replacing RNNs with self-attention and started the effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and has been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head attention and the parameter-free position representation and became the other person involved in nearly every detail. Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and tensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and efficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and implementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating our research.
In any case, if the authors considered their contributions equal, that's good enough for me.
Thanks - wanted to point to this, and indeed should have worded my claim more precisely. And yes, am aware of prior work on attention.
(I need to look it up, but I recall Noam saying publicly that he wouldn’t have agreed to random ordering of contributions if he knew this was going to be this big).
Nope, but it’s not particularly unknown either. It shouldn’t be a surprise; he had remarkable research contributions before and after (separately, he was also an IMO gold medalist).
Some context for people who haven’t followed the full loop: Shazeer was a long-time Google researcher, joined Google in 2000, and was one of the co-authors of “Attention Is All You Need.”
He left Google in 2021 to co-found Character.AI. In 2024, Google brought him and some Character.AI researchers back via a licensing/talent deal with Character.AI (reportedly around $2.7B). He was then made a Gemini co-lead.
At this point is it even pay that’s tempting or is it more about what they get to do? I would assume Google could easily pay them what openAI can, unless as an older company it’s harder for Google to match something really out there
Yeah my current feeling is that once I had double digit millions earning further money would be pretty meaningless to me, and the difference between 'large salary' and 'even larger salary' would be even more meaningless, but who knows maybe it really would change me. I kind of assume people like this are primarily chasing the most interesting/impactful work though.
The problem with this belief is that it implies that all of bigtech is massively overpaying for top talent who would happily stay on for pennies. While bigtech overpaying talent is more plausible than any other bigcorp doing so, it's still rather unlikely.
It gets to the point where what you do is the main question while payment is barely a minor concern way earlier than that point, at least in my experience. You don't need to be in the top AI research tier for that.
OpenAI pays for the earn out he would’ve otherwise received at Google + a new comp package. Made up numbers, if Google still owed him $10M for lasting the full two years, OpenAI can just pay him market rate +$10M.
Yes, but what about the audacity of it? Get paid a lot to join a company but then decide to get up and leave again 2 years later? He just wants to be passed around?
There's a possibility that he lost out in internal political battles, and things weren't going his way. Google is full of battle-hardened political warriors who will do anything (subterfuge, sabotage, etc.) to win battles. It is possible that a guy who just wants to build cool shit would feel like a misfit in such an environment.
I would argue its not the millions though, but rather that sweet rare compute - OpenAI has more of it for his interests than anyone - it is understandable why an exceptional mind would prioritize access to greater capabilities above all else
Right?! Unless you think this move is going to generate general excitement in our lives, it's just another rich guy moving from one high paying job to another.
The Netflix documentary will reveal he was secretly working for Sam Altman the whole time... (Cue diabolical VC-backed evil laugh.)
Google lost three critical years chasing AGI, and got acquired by SpaceX, now a Dyson Sphere startup whose pitch deck is just:
"What if we put a paywall around the Sun?"
I wonder about the motivation to switch teams.
What has Google done wrong? Was he tired?
He could retire, open his own lab, raise capital.
So many opportunities, why go to OpenAI?
Folks talking about the amount of money paid, wasnt he the guy that was acqhired for billions? would OAI pay billions (basically to google) to get him?
[Edit: note that my comment was reparented, it was originally a response to someone claiming Noam was another "Scam Altman". I don't mind the reparenting or the killing of the original subthread, but I feel like this is necessary context to understand this.]
Noam is the real deal, he was pretty legendary within old-time ('00s) Google engineering. Paul Buchheit had a story about interviewing him with the "how to write a spellchecker" question and then him coming up with something better than the state-of-the-art, then basically delivering Google's spell corrector in his first 2-week Noogler project.
Wow, he was using AI to solve problems in 2000 already, that spell corrector being trained on the Web and becoming the first widely used AI tool. Decades ahead.
Just from reading the threads here it seems readily apparent that he then went to start this company that did these bad things. Does not seem confusing at all?
Sorry to bring up the elephant in the room - but could this decision be in part the opportunity to acquire large amounts of stock before a massively inflated IPO?
Google acquired his company in 2024 for $2.7Bn with him taking about 40% of that. I'm quite sure that no matter where he went, any lab or his own start up, he would be fine financially.
I'm sure he was fine financially when he first worked at Google - without leaving to found the startup as well.
But money at that level isn't about being financially secure - to have a roof over your head and food to eat - it's about power.
Money at that level gives you the ability to shape the world in ways others can only dream of - whether that be starting your own company where you can set the values, funding a cure for Malaria, or political lobbying.
Depends on whether the person in question has strong views and a strong belief that they are in the right.
Full disclaimer - I have no insight or knowledge about this particular person - just making the rather obvious and general case that joining OpenAI now at a senior level is likely to generate a serious windfall, and such a windfall is power.
As I said, no idea what motivates this particular person - don't know them at all - the money may be entirely coincidental and it's all about getting stuff done - but he did choose OpenAI rather than somebody like Anthropic....
What does Zionist mean when Israel has existed as a Jewish state for 78 years? I'm genuinely asking because the way the word is used doesn't make sense to me. There aren't similar terms for other countries to just stay the same, like for China to keep being run by the CCP. Every other country is assumed to have ontological inertia except for Israel.
I'm confused, is 78 years a long time? even the US is considered a toddler by empirical terms. zionism wasn't a thing until a minority group had the loudest voice in the room when the allies were discussing what to do with all the european refugees after ww2, and it happened to align well with the brits abandoning their failed colony in the region due to disputes with the locals
here's a quote from wikipedia. it was an utter land grab and an easy way out of responsibility for those in power
> The League of Nations gave Britain mandatory power over Palestine in 1922. British rule and Arab efforts to prevent Jewish migration led to growing violence between Arabs and Jews, causing the British to announce its intention to terminate the Mandate in 1947. The UN General Assembly recommended partitioning Palestine into two states: Arab and Jewish. However, the situation deteriorated into a civil war. The Arabs rejected the Partition Plan, the Jews ostensibly accepted it, declaring the independence of the State of Israel in May 1948 upon the end of the British mandate. Nearby Arab countries invaded Palestine, Israel not only prevailed, but conquered more territory than envisioned by the Partition Plan. During the war, 700,000, or about 80% of all Palestinians fled or were driven out of territory Israel conquered and were not allowed to return, an event known as the Nakba (Arabic for 'catastrophe') to Palestinians. Starting in the late 1940s and continuing for decades, about 850,000 Jews from the Arab world immigrated ("made Aliyah") to Israel.
Yes, this is the important thing to know. I've heard way too many conversations that go back and forth about every act of vengeance in either direction after this, it's all noise. Partition plan started this. But I wouldn't call it an easy way out of responsibility; UK's leaders took a clear and binding position in favor of Zionism.
Also, it was Ottoman territory for hundreds of years up to WWI. I've had friends tell me for some reason about how Palestine was an independent country before... literally wasn't.
You didn't actually answer my question. How does using the word for people who want to create a Jewish state make sense when a Jewish state has existed for 78 years?
One reasonable possibility is they're referring to people like Ben-Gvir who have themselves claimed that Zionism means fighting for Israeli control over more territory like the West Bank. They're the ones calling the shots right now. I don't know whether Zionists 78 years ago would've agreed, it's possible.
To some it still means favoring any existence of a Jewish state. The inertia isn't there because aside from the original partition plan being pushed by the UK, other countries have attacked Israel several times later in ways they would've have withstood without outside support.
"Zionism means fighting for Israeli control over more territory like the West Bank."
Now that is a valid use of the term. I think the problem it that Zionism means so many different things it is nearly useless as a description. It seems more useful as a slur which has become very common in some circles.
"The inertia isn't there"
I'm not sure what you mean. Are you saying Israel could be defeated without US assistance?
I think it's valid to use the word the way that Israel's present leadership is using it.
> Are you saying Israel could be defeated without US assistance?
US and UK, yes. Not just cause of the weapons and money to Israel. After them, the top recipients of US foreign aid in the area are the bordering countries Egypt and Jordan, so that they don't attack.
> I think it's valid to use the word the way that Israel's present leadership is using it.
And how is that?
Israel has a population of 10 million people and a very modern military and nuclear weapons. If it's existence was ever truly threatened things would get VERY ugly.
Because that's what matters. The original Zionists aren't alive to ask what they think. Self-proclaimed Zionists are taking the West Bank and Gaza. In fact they've been kinda doing it for decades under previous governments, but more slowly. If there's some other kind of Zionism around, the most it's doing is complaining, and it's been outvoted.
I have doubts about their ability to self-defend because otherwise we wouldn't be giving so much money, the situation would be stable. Even if they can severely hurt the attackers, it doesn't really matter if the attackers stop at nothing. We just lost a war against Iran despite having full air superiority and killing their leader. And especially if you're considering the scenario where Israel never got Western support, and thus never got those advanced weapons.
Israel left Gaza in 2005.stop telling obvious lies. Hamas attacked Israel on Oct 7 2023 killing at least 800 civilians in an act of incredibly bloodthirsty barbarism, including children, the elderly, and 364 victims attending the Nova music festival. Remember when Hamas paraded the body of that young German woman Shani Louk they killed like a hunting trophy?
IMO people just use the term to mean “pro-Israel” rather than in any reference to the original meaning ("supporter of the idea of a Jewish state"). Which could mean any combination of “pro-American financial support for Israel”, “moral support for Israel in their various military actions”, “opposed to the creation of a Palestinian state”, “a belief that Israel should continue to exist as a Jewish state”, and so on. It's more about the broad political alignment than the specific meaning of the word.
Zionist does have a specific meaning. It means you think the Jewish people have a god-given right to the Palestinian land, and that other creeds and ethnicities should be second class within the Jewish state in Palestine.
A non-zionist Israel would be one where all peoples had the same right, e.g.
"It means you think the Jewish people have a god-given right to the Palestinian land"
It was never actually Palastinian land. It was Jewish land, then Roman land, then Ottoman land, then British land, then Jewish land after Palastinians attacked Israel and lost. At no point were the Palastinians ever a sovereign country and in fact they incredibly foolishly rejected the UN offer for one.
"other creeds and ethnicities should be second class "
Approximately 2.5 to 2.6 million non-Jews live in Israel, comprising about 25% to 26% of the country's total population. This is compared to less than 1% of the population of Gaza being non-muslim.
Surprised to not see more comments on this, especially given the popularity of the Anthropic/Karpathy article. What a win for OpenAI - and what a loss for Google, just 2 years after paying $2.7bn to bring Noam back into the fold. Does not bode well for Gemini long-term... Or could be a signal for how deeply they are leaning into world models.
Trade secrets? Like how to invent a trillion dollar technology and then sit on it for years while others eat your lunch with it? Like how to consistenly release inferior quality models to others despite infinite compute and engineering talent and insane profitability in your legacy businesses?
Not really sure what you're talking about. Apple just licensed Gemini for Siri, Google and their TPU hardware is starting to hit primetime audiences that OpenAI can only dream of.
For those who missed: Gemini coding and agentic capabilities have been lagging the sota models (Opus mostly) since Dec 2026. If you're a co-lead and your model is underperforming there has to be some consequences. I don't know as a fact if this has anything to do with Noam's departure, but work performance is never about past successes.
Don't think it matters in the long run to be honest. The models have no moat, they are becoming a commodity.
Besides that, Google is in a pretty good position, they're not bleeding money on AI like Anthropic/OpenAI, and they own product verticals where they can integrate it. Plus they have a mature ads-model which is what might actually drive a bit of revenue for LLMs.
I think the 'models have no moat' thing is overblown. Only like 3-4 companies in the entire world have cutting edge models, that means there is some kind of moat...
I think when you follow this stuff every day it's easy to lose perspective of the rate of change and these leads seem more profound than they really are when you zoom out a bit.
I'm no super-insider, I only hear industry scuttlebutt like everyone else, but I have about a 95% confidence that the last 18 months has just been about more and better, without any kind of real leap or breakthrough. More hardware, more data, better technique. Well, technique diffuses as people change companies, hardware can be built, and data can be gathered (or stolen!).
From my admittedly outsider perspective, the only years-long moat there is who has the most hardware. If you have the hardware, you can give away the compute to get the data (hello, subsidized subscriptions!). Technique can simply be hired. The only durable, multi-year advantage is the hardware.
So is that a moat? Sure, but it doesn't have a whole lot to do with the leading model companies of the moment. ASML is the real moat, and so it's ASML China is besieging, correctly (IMO) identifying that everything else can be caught up easily enough.
I feel like the models have no moat paradigm died when a single model expanded past the memory of single GPU slices. The moat is hosting the model. Even paying a server host to run a rack of GPUs has immense upstart cost, and then you're still struggling to compete on the add-ons of the things on top of the model (prompts, validation loops, etc). You can only throw so much money at a problem.
yeah, sure, look at anthropic revenue, what is it if not the moat? you can argue for how long but for them good model = the fastest growing company ever.
Grabbing market-share if you have investors that are ready to burn cash infinetely. Find a hot niche, buy a banana 1 USD, sell it for 0.10 USD.
Example: Cursor, they became popular because they were selling ChatGPT unlimited for 20 USD / month.
When they launched, just a reskinned VS Code, "fastest growing AI company"
No coincidence they were bought by SpaceX, who wants to consolidate revenue even if non-sense as long it helps other investors to exit. It shows rapid growth.
Profit is the real moat.
One example: Nvidia. Proprietary tooling, proprietary IP, proprietary hardware, no alternative, expensive.
You don't know what Cursor's game plan was. Maybe acquisition was their plan.
Buying at $1 and selling for $0.1 is still viable as long as they have money in the bank, until they achieve their goals. Most startups start out that way. Even giving away their services for free.
Obviously there will be failures. Doesn't mean they have no moat. Can you say a business with 100 customers and $1000 debt is less viable than one with a single customer and no debt?
And Google has all of those. Custom silicon, more data than anyone else and probably the most comprehensive data collection system, and phones in the hands of 73% of the global smartphone using population to push gemini into to get high volume usage feedback and even more telemetry and data.
I don't think you're honestly accounting for the engineering behind the progress models are making. If it was just a matter of compute on hand and iterating, Meta would be neck and neck with Ant, OAI, and Google, but clearly you've gotta have more.
Noam has a deep expertise in these systems at every level, both algorithmically and at production scale, and knows how to leverage things at different levels.
It's not like Google won't have anyone else that can do what he does, but at the same time, it's an implicit criticism of Google's culture, operations, development, and overall AI program. Shazeer is well past the point where the paycheck is the deciding factor, although I'm certain he is very well paid. Having the freedom to innovate and build free from the corporate fuckery of Google and Facebook is probably more valuable than the pay raise he got with the move, and OAI has the advantage of not having to cope with decades of corporate cruft and inertia. They'll get there - all corporations do - but they're relatively young enough to still be nimble.
I honestly don't think that matters for multiple reasons:
1. There are already multiple "sota" models on the market that compete with only marginal gains between them (OpenAI, Anthropic, Google/Gemini) and some that are catching up (DeepSeek, Qwen,..).
2. The fact that something is a hard engineering problem does not mean it's generating revenue. So while what you said is true, deep expertise is required to push the industry forward, I don't think that is going to matter for the bottom line of these companies. Hence why I think the models don't give a company any 'moat' in a capitalist economy.
I'm curious to know the hype behind the hiring for Karpathy and Noam. In the sense that did oai and anthropic do that for sort of long term and potential new directions (investing in them so they come up with the new transformer). Because it definitely cannot be just a regular filling vacant roles.
Because I think as far as running the existing models and handling whatever nuances, it must be well understood by oai and ANT -- but you don't what you don't know.
Question two: Why are OpenAI spending that money taking talent from Google, who can definitely outspend them for talent, and not Anthropic, who are leading the market and are at least somewhat financially constrained.
Reporting on this seems to indicate that people at Anthropic are significantly more loyal, and that attempts to poach by OpenAI and Meta have been largely unsuccessful.
People seem to have turned down offers that would have netted out more upside for them, so it doesn't seem to just be that. Anthropic seems to lure in the true believers, whereas people are highly skeptical of Sam's motivations these days (particularly after how much safety/alignment has been reportedly cut).
But I'm sure for at least some folks, this is true, given recent valuations.
Allegedly OpenAI is struggling to jump to bigger models and had serious issues in the past (4.5) and also allegedly Shazeer is just the right guy for that. OpenAI is having issues hiring talent as most SF-style people want to go to Anthropic. Shazeer seems more politically aligned with OpenAI. But it's all speculation.
Google has muddied the waters on their Gemini usage statistics as it now powers a big chunk of Search. Depending on how you cut it, Gemini (and Gemini powered products) are probably producing the most output tokens seen by the most human eyeballs by a large margin.
Google at its core is not a dev tools company and it has become evident that is where the money is given the verifiable nature of software. Hixie's reflections on his tenure at Google still ring in my head to this day, though I have never worked there[1].
The people at the helm of Google no longer see the company's identity as something which must be channeled through a product or an experience. Some will point to the DoubleClick acquisition, others will point to Google Reader, or Pichai's ascension. Despite his very short tenure, MBA/McKinsey-brain is a very real phenomenon and it's no mistake that it shaped the "promotion packaged as a product launch" culture that steered Google away from seriously betting on anything that wasn't ads. To quote the signull tweet linked elsewhere in this thread, you can have everything at Google, except for permission.
Most importantly--I don't think there's a single tech product where I can point and say "Google wouldn't do that". You can contrast this with say, other Alphabet companies which don't suffer from this remotely as much. It is VERY clear what Waymo and YouTube are trying to accomplish, and while it frequently makes a ton sense for the companies to share infrastructure and product knowledge, YouTube does an exceptional job on the product side of making it very clear what they would and wouldn't do. They have experimented and shut down experimental features before (is their MOOC functionality still around?), but since it's fairly clear Google specifically is no longer working in service to the mission of providing the world's best digital portal for accessing information, I think it would behoove of them to figure out what their mission is.
I guess this means Google is nowhere close, to even discern
a hint of an AGI? So when Demis Hassabis says AGI...could arrive in just 3 years he has learned the best from Larry Ellison?
In this case, it's not a new thing ... back in 2005 (yes 21 years ago), people talked about the achievements of Noam Shazeer at Google (and Jeff Dean and Sanjay, etc)
I always appreciated Jeff having a level head ... which this article seems to confirm:
Idk, football players actually make a bunch of people happy and entertained. 80% of the United States wishes this tech never existed.
What they're working on is just making peoples jobs, skills obsolete and trying to invent machines that will concentrate the worlds wealth into the hands of the people who own those machines.
Very few people interpret football so much that the actual frontier work of the best players matter. Out of 30 friends I know who like football only 1 of them could explain what’s going on in the field technically. For most people, pro players are replaceable.
Popular entertainment and unique progress of human civilization can’t be really compared either
I'd argue that professional sport is the closest thing to a true meritocracy - doesn't matter who your Dad knows - you ability is there for all to see on the pitch.
And at the team level - if cosy cliques form, again - team performance doesn't lie - hard work, team work and talent is ultimately what delivers results on the pitch.
The other interesting part of professional sport is that the 'workers' have managed to capture more of the value than is traditionally the case - this is precisely because they are so hard to replace.
If you think professional footballers earn too much and are interchangable - feel free to try and get in the team.
Sadly most science and engineering is very capital intensive.
So take this scenario - I'd argue that if you want to make progress in the field of these particular ML models, then you are going to need resources ( compute/data etc ) that is beyond most individuals capability to muster. ie you have to join a company with resources ( or persuade somebody to give you them ).
Right now there is one of those scenarios where capital is chasing talent - and so talent, if they are so inclined, is able to make the most of that.
But in normal times that's typically not the case - most of the time scientists are chasing the capital ( directly or indirectly in the form of a job in a well resourced company ) in order to be able to science, rather than the other way around.
Having the whole world connected to top sports players also costs a lot of money, it doesn’t happen naturally
To become a good scientist you don’t need much classic capital, you need a good environment. And for ML you only need one computer for yourself or you can rent online
There are still big inefficiencies for those who have capital to discover good scientists / engineers. Lots of them are unknown.
But if there are top ones famous it will bring more people to study those fields
This situation is kind of like backend NIL value. His value to OAI isn't just the work he'll do "on the playing field", it's the perceptual value of "OAI just hired the guy Google paid >$2B to get back" right before their IPO.
I doubt that the money had anything to do with it.
I also doubt that the state of the technology at OAI vs. Google had much to do with it, Google is behind no doubt, but the gap is not as far as we know, insurmountable.
I suspect that this is a leadership clash. Noam was working in GDM. GDM somehow went away from coding and RSI into "world models" and that has played out very poorly. Who made that call? Who was still playing politics?
Given this is Noam the list of people that could be pissing him off is very small: Demis, Sergey (?!), a couple of VPs in GDM.
You can't force someone to keep taking your money (that's indentured servitude), you can only incentivize them to stay with increasing amounts of money. Google almost certainly did do that. Probably by vesting his hiring bonus over 2-3 years.
OpenAI is in a unique position right now to grant pre-IPO options (probably in the form of RSUs). And they wanted him badly enough to grant the extra options necessary to effectively 'buy out' whatever unvested Google bonus he's walking away from.
It gives some context on the contributions of each of the authors. About Shazeer, from the article:
Shazeer’s joining the group was critical. “These theoretical or intuitive mechanisms, like self-attention, always require very careful implementation, often by a small number of experienced ‘magicians,’ to even show any signs of life,” says Uszkoreit. Shazeer began to work his sorcery right away. He decided to write his own version of the transformer team’s code. “I took the basic idea and made the thing up myself,” he says. Occasionally he asked Kaiser questions, but mostly, he says, he “just acted on it for a while and came back and said, ‘Look, it works.’” Using what team members would later describe with words like “magic” and “alchemy” and “bells and whistles,” he had taken the system to a new level.