For AI researchers pursuing AGI, this variance between distributions is arguably even worse than the distribution between samples - there's no past data whatsoever to build estimates, it's all vibes.
You can argue the distribution is hard to pin down (hence my note on risk), but let’s not pretend there’s zero precedent.
If it turns out to be another winter at least it will have been a fucking blizzard.
But the distribution for individual researcher salaries really is pure guesswork. How does the datapoint of "Attention Is All You Need?" fit in to this distribution? The authors had very comfortable Google salaries but certainly not 9-figure contracts. And OpenAI and Anthropic (along with NVIDIA's elevated valuation) are founded on their work.
When Attention is All You Need was published, the market as it stands didn't exist. It's like comparing the pre-Jordan NBA to post. Same game, different league.
How much revenue does Google make in a day? £700m+.
None of these models are operating in a vacuum.
“Our Rock Stars Aren't Like Your Rock Stars”
Like I definitely think it is better for society if the economic forces are incentivizing pursuit of knowledge more than pursuit of pure entertainment[0]. But I think we also need to be a bit careful here. You need some celebrities to be the embodiment of an idea but the distribution can be too sharp and undermine, what I think we both agree on is, the goal.
Yeah, I think, on average, a $100M researcher is generating more net good for a society (and world) than a $100M sports player or actor. Maybe not in every instance, but I feel pretty confident about this on average. But at the same time, do we get more with one $100M researcher or 100 $1M researchers? It's important to recognize that we're talking about such large sums of money that at any of these levels people would be living in extreme luxury. Even in SV the per capita income is <$150k/yr, while the median income is medium income is like half that. You'd be easily in the top 1%. (The top 10% for San Jose is $275k/yr)
I think we also need to be a bit careful in recognizing how motivation can misalign incentives and goals. Is the money encouraging more to do research and push humanity's knowledge forward? Or is the money now just another means for people that just want money to exploit, who have no interest in advancing humanity's knowledge? Obviously it is a lot more complicated and both are happening but I think it is worth recognizing that if things shift towards the latter than they actually make it harder to achieve the original goals.
So on paper, I'm 100% with you. But I'm not exactly sure the paper is matching reality.
[0] To be clear, I don't think entertainment has no value. It has a lot and it plays a critical role in society.
For whatever reason, remuneration seems more concentrated than fundamentals. I don't begrudge those involved their good luck, though: I've had more than my fair share of good luck in my life, it wouldn't be me with the standing to complain.
> Katalin Karikó was thought to be working in some backwater, on this "mRNA" thing, that could barely get published
There's a ton of examples like this, and it is quite common in Nobel level work. You don't make breakthroughs by maintaining the status quo. Unfortunately that means to do great things you can't just "play it safe"Locking up more of the world's information behind their login wall, or increase their ad sales slightly is not enough to make that kind of money. We can only speculate, of course, but at the same time I think the general idea is pretty clear: AI will soon have a lot of power, and control over that power is thought to be valuable.
The bit about "building great things" certainly rings true. Just not in the same way artists or scientists do.
People forget Kerr was a bad GM
Don’t get me wrong, they are smart people - but so are thousands of other researchers you find in academia etc. - difference here is scale of the operation.
Even if it’s 1% at the scale you’re talking that’s 1B to the company. So still worth it.
Wild.
You bring up the only relevant data point at the end, as a throw in. Nobody outside of academia cares about your PhD and work history if you have a startup that is impressive to them. That's the only reason he's being paid.
"Our key innovation is a new collection of datasets called PixMo that includes a novel highly-detailed image caption dataset collected entirely from human annotators using speech-based descriptions, and a diverse mixture of fine-tuning datasets that enable new capabilities. Notably, PixMo includes innovative 2D pointing data that enables Molmo to answer questions not just using natural language but also using non verbal cues. We believe this opens up important future directions for VLMs enabling agents to interact in virtual and physical worlds. The success of our approach relies on careful choices for the model architecture details, a well-tuned training pipeline, and most critically the quality of our newly collected datasets, all of which we have released."
This is a solid engineering project with a research component - they collected some data that ended up being quite useful when combined with pre-existing tech. But this is not rocket science and not a unique insight. And I don't want to devalue the importance of solid engineering work, but you normally don't get paid as much for non-unique engineering expertise. This by no means sounds unique to me. This seem like a good senior-staff research eng project in a big tech company these days. You don't get paid 250M for that kind of work. I know very talented people who do this kind of work in big tech, and from what I can tell, many of them appear to have much more fundamental insight and experience, and led larger teams of engineers, and their comp does not surpass 1-2M tops (taking a very generous upper bound).
> given what even a resource-constrained DeepSeek did to them.
I think a lot of people have a grave misunderstanding of DeepSeek. The conversation is usually framed comparing to OpenAI. But this would be like comparing how much it cost to make the first iPhone (the literal first working one, not how much each Gen 1 iPhone cost to make) with the cost to make any smartphone a few years later. It's a lot easier and cheaper to make something when you have an example in hand. Just like it is a lot easier to learn Calculus than it is to invent calculus.Which that framing weirdly undermines DeepSeek's own accomplishments. They did do some impressive stuff. But that's much more technical and less exciting of a story (at least to the average person. It definitely is exciting to other AI researchers).
If you do not believe this narrative, then your .com era comment is a pretty good analysis.
> There is a group of wealthy individuals who have bought in to the idea that the singularity is months away.
My question is "how many months need to pass until they realize it isn't months away?"What, it used to be 2025? Then 2027? Now 2030? I know these are not all the same people but there are trends of to keep pushing it back. I guess Elon has been saying full self-driving is a year away since 2016 so maybe this belief can sustain itself for quite some time.
So my second question is: does the expectation of achievements being so close lengthen the time to make such achievements?
I don't think it is insane to think it could. If you think it is really close you'd underestimate the size of certain problems. Claim people are making mountains out of molehills. So you put efforts elsewhere, only to find that those things weren't molehills after all.
Predictions are hard and I think a lot of people confuse critiques with lack of motivation. Some people do find flaws and use them as excuses to claim everything is fruitless. But I think most people that find flaws are doing so in an effort to actually push things forward. I mean isn't that the job of any engineer or scientist? You can't solve problems if you can't identify problems. Triaging and prioritizing problems is a whole other mess, but it is harder to do when you're working at the edge of known knowledge. Little details are often not so little.
It's going to persist until shareholders punish them for it. My guess is it's going to be some near-random-trigger, such as a little-known AI company declaring bankruptcy, but becoming widely reported. Suddenly, investing in AI with no roadmap to profitability will become unfashionable, budget cuts, down-rounds, bankruptcies and consolidation will follow. But there's no telling when this will be, as there's elite convergence to keep the hype going for now.
Telco capex was $100 billion at the peak of the IT bubble, give or take. There's going to be $400 billion investments in AI in 2025.
they know it may be or not gonna happen because months its ridiculous, but they still need to do it anyway since if you not gonna ride it, you are gonna miss the wave
stock market has not been rational since??? forever??? like stop pumping and dumping happen all the time
While there's a lot of money going towards research, there's less than there was years ago. There's been a shift towards engineering research and ML Engineer hiring. Fewer positions for lower level research than there were just a few years ago. I'm not saying don't do the higher level research, just that it seems weird to not do the lower level when the gap is so small.
I really suspect that the winner is going to be the one that isn't putting speed above all else. Like you said, first to market isn't everything. But if first to market is all the matters then you're also more likely to just be responding to noise in the system. The noisy signal of figuring out what that market is in the first place. It's really easy to get off track with that and lose sight of the actual directions you need to pursue.
Remember capsule networks?
those figures is payment for access to know-how knowledge developed by competitors, those people are not necessary "extremely smart".
They're not high because of performance/results alone.
Yes, the figures are nuts. But compare them to F1 or soccer salaries for top athletes. A single big name can drive billions in that context at least, and much more in the context of AI. $50M-$100M/year, particularly when some or most is stock, is rational.