> better products downstream
Maybe if they somehow fixed hallucinations and kept the secret sauce to themselves, I could see them being worth that much, but all the top labs seem to have given up on that problem.
Anthropic could pivot to a UI and hosting for models (potentially with some or no proprietary models) and still be worth $60B.
Anthropic's team plans are actually pretty pleasant to use. It does seem strange/funny though, that a significant part of the evaluation choice from a business perspective is that Anthropic is more trustworthy than OpenAI.
In private markets, especially with the recent trend of selling a tiny portion of the company at a massive price, the valuation represents something much closer to the maximum that any investor in the world thinks the company is valued at.
Amazon can "buy" $2B worth of Anthropic to guarantee $2B of spending on AWS - to report that as growth under AWS in their earnings - to juice their stock price.
They also get to report that their investment in the previous round is up massively.
This is all before accounting for the preference stack, which makes multiplying a Series F per-share price (itself derived from dividing compute time by some magic number) by employee common stock a bit silly.
That’s still something, especially right at the money!
Some series also have various blocking, dividend and other rights.
1. We create near AGI
2. We create ASI (Artificial Super Intelligence)
In the first scenario, an investment in any AI model company that does not own its own compute is like buying a tar pit instead of an oil well. In other words, this future has AI like a commodity and the entity who wins is the one who can produce it at the lowest cost.
In the second scenario, an investment in any AI model company is like buying a lottery ticket. If that company is the first to create ASI then you've won the game as we know it.
I think the minor possibility of the second scenario makes this a good investment. But it definitely feels like all or nothing instead of sustainable business.
ASI doesn't mean it can magically reverse-infer a digital process from an existing hodge-podge of automation and manual steps.
ASI doesn't mean there are instantly enough telerobotics or sensors to physically automate all processes.
Similarly, even ASI will by definition be non-deterministic and make mistakes. So processes will have to be reengineered to be compatible with it.
Even that is small minded frankly. The economy would not survive a super intelligence.
Can the smartest developer you know build a project correctly without any requirements?
There's a lot of sausage making behind what to build, how to integrate it with other things, who needs to be told what on which other teams, etc.
Even pie in the sky breakthrough ASI isn't going to be able to do all of that, Day 0.
And if we're using a tautology to define ASI as something that can do that on Day 0, then I'd point out that in the entire history of technology there hasn't been a single advancement that wasn't subsequently refined and improved.
Except maybe fire.
If not, how much smarter do you think someone would need to be to do so?
ASI ridiculousness starts from defining it as a do-anything machine. Everything has limits.
Like… I’ve had a TODO that I’ve expanded and worked on for a few years for a personal project. If I had a tireless bot of my own intelligence, it should only have taken about an hour. Human limitations are immense barriers.
It's tricky since the future of AI isn't something anyone can really prove / disprove with hard facts. Doomers will say that the rate of improvement will slow down, and anti-doomers will say it won't.
My personal believe is that with enough compute, anything is possible. And our current rate of progress in both compute and LLM improvement has left Doomers with shaky ground to discount the eventuality of an AGI being developed. This just leaves ASI as a true question mark in my mind.
Are we seeing the same progress? GPT-4 was released in March 2023, that's almost two years. Tools are much better but where is the vast improvement?
I still Google things I want to know and skip the AI part.
My Google use is down significantly. And I mostly reach for it when I am looking for current information that LLMs do not yet have training data for. However, this is becoming less of an issue as of late. DeepSeek for example has a lot of current data.
Dunno, we're already at ridiculous amounts of compute and progress has slowed, a lot. I think we need another technological breakthrough, a change in technique, something. LLMs don't seem to be capable of actually learning in the way humans do, just being trained on data, of which we've reached the limit.
Endless growth and technological improvement isn't the only option, and seems to me like the least likely. The other option means that there will be a peak somewhere.
This took me down a memory lane:
- Dragon Dictate speed recognition improvement curve in the mid-90s would have led to today's Siri sometime around 1999.
- The first couple of years of Siri & Alexa updates...
- Robots in the '80s led us to believe that home robots would be more or less ubiquitous by now. (Beyond floor cleaners.)
- CMU winning the DARPA Urban challenge for autonomous vehicles was a big fake-out in terms of when AVs would actually land.
Most of the benefits of computing come from relatively small improvements, continuously made over many years & decades. 2-4 years is not enough time to really extrapolate in any computing domain.
> with enough compute
"enough" here could be something that is only measurable on the Kardashev scale.
Why do you assume that AI will become a commodity that is only metered by access to compute?
Right now (since June 2024), Anthropic is ahead of the field in quality of their product, especially when it comes to programming. Even if O1/O3 beat them on benchmarks, they are still nowhere near when normalized for compute needs.
Can they sustain this? I don't know, but in the end this is very similar to known software or even SAAS business models. They are also in a somewhat synergetic relationship to Amazon.
Did office software ever become commoditized? Google and many more tried hard, but there is still the same company in the lead that was in the 90ies.
there is this sentiment on internet, but in my personal experience, GPT4 hallucinate APIs and usage examples way less, and after trying to get Claude working, I switched to GPT as the first step in my coding workflow.
All AI stocks come crashing down from fantasy amounts of money to what they’re actually worth as they did with every previous tech hype until people find a new thing to throw their money at.
do you read reddit? lots of people see through the facade now. the only people excited are shills
It's pretty simple to go into threads about breaking events several years later and see them loaded with confidently incorrect statements dominating the discourse.
One of my favorite examples was the $34 IPO stock offer to moderators and redditors, which was almost universally bashed (currently trading over 170).
https://www.reddit.com/r/investing/comments/1b0n2eo/reddit_i...
Reddit just turned a profit after nearly 20 years in business.
Its whole use maybe upended by AI and bots, so even if its profitable now hard to know if it will continue.
I don’t know what drives it. I think part of it is that zero commission stock trades have many more small players making trades and stocks are priced on the margin.
Certainly not debating that this isn't possible, ofc. As someone whose spent the past year+ working full time on the philosophy of this technology I think you're going against a pretty clear scientific consensus among AI experts, but perhaps you have your reasons.
I think we already have AGI anyway, so I'm either a loon or a pedant ;)
Overhyped (and more specifically, overvalued) is a different question though. I think that most people working in the AI field have a pretty good monetary reason to say that AGI and ASI are just around the corner, but I have yet to see proof of any of it being achieved in any way whatsoever.
Looking at everything we have so far, LLMs are still only token prediction and neural nets can only do what they’ve been trained to do. The datasets may be bigger, the computing power and efficiency may be increasing and we’re building abstractions that make longer chains of „thought“ possible, but at the end of the day this is still the same technology with the same restrictions that it’s always had and I don’t see that changing.
I’m honestly betting on robotics. Tokenising words is intuitive. But the parameter space for tokenised physical inputs is both better understood and less solved for.
3. We create incrementally better versions of generative AI models which are incrementally better at making predictions based on their training sets, are incrementally more efficient and cheaper to run, and are incrementally better integrated into products and interfaces.
In my opinion, this seems to be the more likely than some of the other wilder scenarios being predicted.
Thats why I like the idea of openrouter so much. Next token prediction is a perfect place to run a marketplace and compete on price / speed.
It's hard for me to see a future where the long term winners aren't just the chipmakers and energy companies.
With the o3 benchmarks its becoming apparent the primary thing from keeping current generation of models from getting smarter is more processing power. If somehow chips got 10x faster tomorrow, it still wouldn't be fast enough. Even at 1000x faster than current performance, those $3k o3 queries now cost $3 which are still too much.
If you invest in a data center, any amount of billions you put in will not future proof it because faster chips are always around the corner and you will be at the mercy of suppliers.
If you invest in a model, even if you invested billions in 1-10 years people will be able to run equivalently powerful models on consumer hardware.
I'm loving the competitive system LLMs created by having a unified API interface allowing you to swap out models and providers with single lines of code.
That is kinda startup funding in general isn't it?
You can have as much intelligence in a model/system as you want, and it can well be ASI, but as long as this intelligence doesn't have the resources it needs to run at its full potential, or at least at one higher than the competition, you're still not over the hill.
Ultimately those companies or countries which will have the most resources available for an ASI to shine will be the ones which win.
Once ASI has figured out how to obtain energy (and ICs?) "for free", or at least cheaper than the competition, it will have won.
More alarmingly, such a thing would be more dangerous than a nuclear weapon, and might reasonably merit a nuclear first strike.
Capital has been an economic multiplier of human labor. AGI is more like flooding the market with infinite labor. There will be blood in the streets, at best.
A tractor was one of many things that humans have invented that could be considered a labor multiplier and put many people out of work.
AGI is just more people. It’s infinite people in the labor market driving wages down to zero.
The new “reasoning” models get very marginal improvements in output for huge increase in token count and energy use.
None of this is sustainable, and eventually, and soon, crops will start to fail en masse due to climate disasters.
Then the real fun starts.
Right now “AI” is more harmful than helpful on a species level balance sheet. These are real problems, today.
The last valuation was at $157 billion- Anthropic is valued at 1/3 of OpenAI but has 1/10th of the market share....
Also, Anthropic's Sonnet 3.5 seems to be widely preferred as a developer tool, even over OpenAI's newer GPT-o1, and developer use is one of the current leading use cases for AI.
Which says Anthropic is about 1/2 the API revenue of OpenAI and growing fast. But OpenAI is actually 5x revenue overall and 18x the chat product revenue. (This is from Oct, not sure how much would have changed).
Especially considering I pay for Claude, the beer, and my business _thru Stripe_!
P.S. I'm so glad you're able to derive some joy from these new technologies, but I would also offer a soft suggestion to watch Season 3 of Westworld. It's probably not as good aesthetically as the previous two, but it's also pretty separated and deals throughout with the concept of AI therapists/friends, and how they might be a short-term comfort but a long-term threat to our individuality. Obviously a chat here and there with current LLMs is nowhere near that yet, but thought you might find it interesting!
On topic though, I wholeheartedly agree that this valuation seems rather... unrealistic. I do think "hype" is how we currently handle a lot of valuations, for the worse in my opinion.
with all due respect, how can you take yourself seriously doing this? I tried to use an LLM as a "cheap therapist" exactly one time and it felt so phony I quit almost instantly, after about four messages.
The bot pretends to feel compassion for you! How does that not induce rage? It does for me. False empathy is way worse than nothing.
And on top of it, you're talking to a hosted LLM! I hope you are not divulging anything personal to the random third party you're sending your thoughts to..
This stuff is going to be such a huge boon to authoritarian governments.
Feel free to pass over useful tech, but no need to disparage others for not wearing your tinfoil hat.
In light of all that it's hardly "chicken littling" to assume that hosted AI chat bots can't be trusted with intimate and personal details. The companies running them should be seen as inherently untrustworthy.
- build structure into conversation
- identify root causes of issues
- provide advise about behavior/thinking changes which could mitigate root causes
You’re not wrong, but you’re not right either. For whatever it’s worth, I absolutely plan on self hosting (nvidia Digit!) my future conversation partner. But nothing about me is terribly private. I love my wife and my sons and sometimes I wonder if I’ll be forgotten or if we’re about to give birth to the overman. It’s nothing I wouldn’t tell a stranger.
Lastly, no therapist I’ve ever met can talk about my favorite authors with me for hours, ad-hoc, on-demand, for pennys a day. All my real friends are sick of Dostoyevsky, but not Claude!
This is how I feel and it is so rare and refreshing to see someone else say it.
It appears to me that most people are very private and will go to great lengths to protect their privacy.
If privacy wasn't of particular concern, this is as near an "absolute good" idea as one can forge.
You're not investing hoping that they turn into a big business with a nice return. You're investing because you assume the value will either be zero or infinity.
If they achieve AGI first, then the valuation you invested in doesn't matter because the value will basically be infinite (or will completely change society in a way that money won't matter anymore).
If someone else achieves AGI first, the value is basically zero.
And if AGI isn't achieved, well, there probably won't be any exit. But if there is, it's a nice bonus if it still has any value.
His thesis that AI is just distilled omega-capitalism is true, very plainly.
Say you're some OpenAI employees. You know a lot about AI, you've got the resume, you've got some buzz, and you want your own successful startup. How do you make sure that it gets maximum preferential treatment and first dibs on all the data and GPU? By making sure the big dogs are all going to get super rich off of your success.
So they got billions in investment from Google, even more billions in investments from Amazon, half a billion from FTX (whoops), then some VCs for additional shmoozing power, and you're good to go. It helps to have the technical chops and a good product to distinguish yourself, but at that point, Amazon and Google are both going to go out of their way to shove your AI into everything, so having something to contribute is practically just a nice bonus if you can manage it.
The top 1% in US own 40% of wealth, and top 10% own 80% of it. We are beyond Pareto ratio of 20:80.
I've created some new ideas: Yes, maybe all I did was "read what other's had done and took the next step", but that step seemed novel to me and, thus, just NOT in the "essays" or "what other's had done". Uh, just how the novelty happened does not seem to be in any of the essays or "what other's had done"?
Okay, maybe a two-step approach: (1) Make wild guesses. (2) Run experiments and test the guesses. But is current AI doing either of (1) and (2)? Right, for some board games can do both (1) and (2). Test the guesses with the content of the essays or "others have done"?
Here's a simple example: At one point the FedEx BOD wanted some revenue projections, uh, seriously wanted as in else "pull the funding". People had hopes, wishes, but nothing that sounded objective. Soooo, I noticed, guessed (1) growth would be mostly from the happy existing customers (2) influencing customers to be. The influencing would be a customer to be receiving via FedEx a package from a happy customer. So what? Okay, for time t let y(t) be the revenue at time t. Let b be the total size of the market, i.e., the revenue when do have all the target customers. Then at time t, the growth rate would be proportional to both (a) the number of current current customers and, thus, proportional also to y(t) and (b) proportional to the number of customers to be, and, thus, to (b - y(t)). So for some constant of proportionality k, we have that the growth rate
d/dt y(t) = y'(t) = k y(t) (b - y(t))
which has a simple closed form solution. Then for any k > 0, can do some arithmetic and find y(t) for any t > 0 and draw a graph. Do this and pick a k that yields a plausible, reasonable graph, and present that to the BOD. It worked, i.e., pleased the BOD which did not "pull the funding".
This work was 100% mine. Lot's of people in the office worked on the problem, but none of then had any ideas as good as mine -- i.e., could have them all write the "essays", process those, and still not come up with the little differential equation, its closed form solution, or a reasonable k.
It seems to me that the essays, what others have done as training data just does not have or have a path to work that is new, correct, significant. Uh, can we train the AI in how to guess and test (beyond board games), how to start with a BOD request and, a description of the why and how of business growth, some calculus and get an answer, take epicycles and come up with F = ma, Tesla's experiments, Stokes formula, and get Maxwell's equations, make a wild guess and propose the Michelson-Morley experiment, get E = mc^2, use the inverse and implicit function theorems, Riemann's work on manifolds, and get general relativity, solve Fermat's last theorem, make real progress on the Riemann hypothesis??? Uh, in short, we need a idea not in the training data? Soooo, need to train the AI to have ideas? Ideas are just using what's in the essays to make connections in a graph and then exploring the graph until get a path to an answer?? How do such training? Can process existing text yield such training data?