jononor

They could be be playing into such a narrative, even though they have different motivations. Certainly Altman has had a few moments where he is saying "this stuff is getting so good it is society-scale dangerous". His motivations are to create FOMO with investors and users, and to try to form the regulatory landscape in their favor.

I do not think they are aiming to be cashflow positive now. That might not be possible. Though if it is within range they might want to go for it. Because the stamina needed to win this race is going to be immense. Especially since OpenAI is going hard for scaling up via investor funding, and Google can afford to loose/invest a couple billions annually by diverting from their main revenue sources.

A realistic business plan would be to burn cash for many years (potentially more than a decade), and bank on being able to decrease costs and increase revenue over that time. Investors will be funding that journey.

So it is way too early to tell whether the business plan is unsustainable. For sure the unit economics are going to be different in 5 and 10 years.

Right now is very tough though- since it is basically all early adopter power user types, which spend a lot of compute. Later one probably can expect more casuals, maybe even a significant amount of "gym users" that pay but basically never uses the service. Though OpenAI is currently stronger for casuals, I suspect.

Over the next decade, hardware costs will go down a lot. But they have go find a way to stay alive (and competitive) until then.

It was never unlimited. The 5 hourly limit was there from before.

That just gives your team another thing to do, trying to police a ToS clause. And one would have to define abuse somehow, which is quite tricky. There are a lot of guite legitimate but expensive ways to use such a general thing as LLM compute.

Investors love MRR/ARR, so I do not think we will see that as the main option anytime soon. That said, you can use the Claude API to get usage-based billing.

Would have loved to get some more insights. Cost estimates, before and after, for example. But also if any architectural changes where needed, or what kind of other challenges and learnings they got from the migration.

ChatGPT and Claude do the same. And I have noticed that model performance can often degrade a lot before such a hard limit. So even when not hitting the hard limit, splitting out to a new session can be useful. Context management is the new prompt engineering...

Agreed. I chose gevent over asyncio for a backend in 2019, and we still using it. Works pretty well. No plans to phase out gevent just yet.

Though the community has clearly centered on asyncio by now. So if I were to start a new backend today, it would reluctantly be asyncio. Unfortunately...

And spaces could be variable in the number of "sub-spaces" used. Each could be say 1/8 of a regular space.

Some percentage will not comply. Either in bad faith or just lazyness/incompetence/accident/whatever. But as long as that number is relatively low and a decent chunk pays "their dues", it is not really a problem.

Why is "AI" (current LLM based systems) a danger on the level comparable to nukes? Not saying that it is not, just would like to understand your reasoning.

I agree with your reasoning. But the conclusion seems to be throwing the baby out with the bathwater?

The same line of thought can be used for any (new) tool, say a calculator, a computer or the internet. Shouldn't we try to find responsible ways of adopting LLMs, that empower the majority?

I like the idea. But it is missing something to actually get money out of companies, I think? Because even when there are people in a company that are nominally willing to pay, there is usually so much friction/hassle to actually get money out of a company - that it most often ends up not happening for open source. Unless there is something that "forces" them.

One nice thing about GitHub sponsorship is that there is only one bill for the sponsor, and one can support NN projects/creators there. I think it is even bundled with the regular Github invoice?

I would only say that they have become maintainers if they consistently do so, over time. Including helping out in areas which is not directly useful to them. And thinking about the whole picture, not just individual features and bugs. And also putting in the time when it is needed, even though other obligations are pulling at them. Of course it is somewhat of a continiuum. And it always starts with being a contributor. So that is on the right path.

Another very good way to lie is to set up the framing such that any interpretation of any fact skews in your desired direction. Including which things are to be considered important/relevant, what kind of argument is considered valid/not. Done well, people might not even pick up that there is lying/misdirection involved. Rig the game.

I concur. This post of one being welcoming, accepting and encouraging is appreciated. What is the opposite of "gatekeeping"?

That expert systems were not financially viable in the 80-90ies does not mean that LLM-as-modern-expert-systems cannot be now. The addressable market for such solutions has expanded enormously since then, thanks to PCs, servers, and mobile phones becoming ubiquitous. The opportunity is likely 1000x (or more) what is was back in those days - which shifts the equation a lot as to what is viable or not.

BTW, have you tried out your challenges with a LLM? I expect them to be tricky to answer direct. But the systems are getting quite good at code synthesis which seems suitable. And I even see some MCP implementations for using constraint solvers as tools.

Superhuman capability within the tasks that are well represented in the dataset, yes. If one takes the view that intelligence is the ability to solve novel problems (ref F. Chollet), then the amount of data alone might not take us to superhuman intelligence. At least without new breakthroughs in the construction of models or systems.

R1 managed to replicate a model on the level of one one they had access to. But as far as I know they did not improve on its predictive performance? They did improve in inference time, but that is another thing. The ability to replicate a model is well demonstrated and quite common practice for some years already, see teacher-student distillation.

If OpenAI would publish the models before the competition, then one could verify that they were not tinkered with. Assuming that there exists a way for them to prove that a model is the same, at least. Since the weights are not open, the most basic approach is void.

This is solvable? They could post a hash of the solutions publicly, and reveal the contents of the solution after one week.

It would leave a lots of general purpose GPU-based compute. That is useful and enduring infrastructure? These things are used for many scientific and engineering problems - including medicine, climate modeling, material science, neuroscience, etc

A) We are out of the Internet-scale-for-free data. Of course the companies deploying LLM based systems at massive scale are of course ingesting a lot of human data from their users, that they are seeking to use to further improve their models.

B) Has learning though "self-play" (like with AlphaZero etc) been demonstrated working for improving LLMs? What is the latest key research on this?

What exactly do you mean by replace? Replace in particular roles (causing the humans to shift into new roles), or replace as in nothing useful left for humans to do? Those are two very different trajectories.

So far in the industrial revolution we have been experiencing the first. Waves of automation that displaces workers in one role, and then new roles open up for humans (often involving the machines). I believe that for the foreseeable future human + AI tools (symbiosis) will be a much stronger than AI alone. Of course it will tend towards more and more AI per human over time. The same way modern manufacturing is "machine tending", where a few workers might supervise a factory that outputs what would have taken thousands of workers in the past. If the thing being produced is has unrealized demand, then we would expect that being able to do it more effectively would just mean more production, possibly to the extent that the same or even more humans are needed to produce it. So the real questions are - which things does humanity have a lot more demand for (2-1000x) - and can be made more effective with AI in the loop / on the team. And the flip side, which things are we near max demand for, and can also be made more effective with AI. Jobs in those areas are going to be decimated, move away as quickly as possible.

What would the "horse" equivalents be in this AI tech shift? Legit question - getting reasonable answers here might help us understand where we are going.

For sure it is not computers. Or software systems. We will for sure have way more of those. "Code" is more unclear, there will for sure be more of it - but for sure less human-hours per line of code. So the roles there will change quite a bit, though unclear exactly how. More focus on systems, business rules, specification, I suspect. Translations (between human languages) seems like a thing that is highly risky. It seems that it could go the way of the horse, automated translations being 99% of the volume eventually?

Ahahaha, that is legit a nice framing and "business model" for a Patreon in the AI space. Drinking from the firehose and washpanning it, so that not everyone has to do the same.

Yeah, spending a few hours trying things out will be worth much more than spending hours reading about it.

It is pretty easy and convenient to write extensions using pybind11, including passing numpy arrays. It takes 10 lines in setup.py and around 10 lines in a .cpp file, run setup.py build_ext to build it. Not quite the convenience of inline - but in practice pretty nice. My only nit in that compile time is around 3 seconds on my machine.

The Internet was also very expensive in its infancy. Dialup charged by the minute for mere kilobytes. The cost per MB dropped by a factor 1000x over the course of 30 years. It took billions in investments, and millions of people working on it to make it happen. Give LLLms a couple of decades, and the price for a given capability will have increased by 1-4 orders of magnitude.

Just as one example, MicroPython is deployed in several satellites (cubesats). There are some conference talks and podcasts about it.