I used to work in web development and I still do some programming as a hobby, but I'm currently a fifth grade teacher (because I love working with kids).
My email is my username and gmail. As long as you're not a bot and you're not trying to sell me something, I'd really love to hear from you!
I was going to reply that Anthropic, which supposedly is the most capacity constrained of the leading AI labs, still provides access to models as old as Opus 3.
Biologists celebrate the discovery of new species of fruit fly hidden deep in the Amazon rainforest. Astronomers celebrate the discovery of new giant rocks located zillions of light years away. Neither of these things is immediately “useful” to the world, although they may turn out to be enormously beneficial in ways we can’t immediately predict. To me, these fields also feel central to the human experience—discovering new types of life, or learning more about our place in the universe. I don’t think a mathematical proof is any different.
I suspect part of the problem is that the author is fighting the system prompt, which gives Claude instructions to help it avoid filling up its context window.
So the author thinks he's giving Claude this instruction:
> You must re-read CLAUDE2.md, even if you've already read it before.
But the actual instruction is closer to:
> Do not re-read files you have already read. You must re-read CLAUDE2.md, even if you've already read it before.
So Claude has conflicting instructions. Is it any surprise that it tries to thread the needle by re-reading the minimal amount of CLAUDE2.md necessary? It's just doing its best to satisfy both masters!
I've never understood why people make fun of Rust code that has lots of unsafe blocks. Obviously, the goal is to reduce those blocks, but consider also the number of safe blocks!
...I think it really is irrelevant, isn't it? The LLM gets words as tokens, not strings of letters. If you asked me how many of the letter s is in Mississippi, but said I'm not allowed to spell out the word in my head and count the letters, I don't think I could do it.
This isn't a great analogy, because part of the challenge would be preventing myself from picturing the spelling in my head. But my point is, the AI is not getting the words as letters. The correct solution is tool use.
> In the story they spend months to build the MVP and people don't like it.
Oh, that was another problem! The story says:
>> In practice it doesn’t work very well, but it’s good enough for an MVP.
No, that's quite clearly not good enough for an MVP! If your product's core functionality—baking stuff, in this case—only works one third of the time, then your product is not viable! It is maybe a proof of concept, but certainly not a product.†
So the company never should have released an MVP, they should have kept working on their reliability problems. Which is when the engineer should have realized "alright, this isn't going to work," and the whole company needs to pivot or close up shop.
Alternately, if they ship a real MVP that can produce perfect cakes with reasonable reliability, and the customers don't like it, that would also be time to change plans.
What happened in the story is they shipped a product that obviously doesn't work, customers dislike the fact that it doesn't work, and the obvious conclusion is "well, they didn't like it because it doesn't work yet," which is true but also obvious. You didn't learn anything.
This doesn't really answer your question and I don't think a definitive answer exists (and I've never worked in this space). But, obviously, this company should have stopped sooner than they did.
---
† The only exception I can see is if you can come up with a potential customer for which "an oven that bakes correctly one third of the time" would still meet their needs. Under the scenario in the story, I don't think this exists. For, say, an AI coding agent that can run automated tests and throw away bad results, a one third success rate may still be useful in some scenarios.
> The engineer realizes something: building an algorithm that calculates baking time for cakes, pizzas, and bread is quite a bit more complex than it looked. Every dough is its own universe. They need to hire more engineers.
Why did the engineer "who spends all day talking and arguing about ovens" not realize this sooner? Sure, "it is difficult to get a man to understand something when his salary depends on his not understanding it", but the engineer's salary wasn't great anyway; the real goal was to build "the oven of his dreams". To do that, he very much needed to understand the algorithmic complexity involved.
What I assume happened is the engineer wasn't sure whether the idea could work, and the only way to find out was to try. Well, he tried, and he found out. Oh well.
Isn't this how VC is supposed to work? Ten startups try ten ambitious ideas. Nine fail, one succeeds. The one that succeeds does well enough to make up for the nine failures. And so it goes. There was nothing wrong with the nine founders who failed. They were just unlucky, and they can try again.
I think what went wrong in the story is very simple. The company didn't "fail fast".
> They do have at least one development studio there, Nintendo Software Technology.
Oh, interesting! The list of games isn't particularly impressive though, a lot of ports and remakes. Compare that to also-US-based Retro Studios, which isn't considered under Nintendo of America from what I understand.
Oh come now. Fable was available for less than a week before it was pulled, not enough time to build a business critical function. The government isn't going to pull a model that has been out for a substantial length of time. Or do you also avoid using US-developed encryption?
(Okay, I can't predict what crazy crap this particular administration is going to do, but that goes for anything, well beyond AI. I think it's about as likely that they would restrict access to Opus as restrict exporting iPhones, or bomb Greenland, or whatever.)
> But one thing scares me a little. What if LLMs slowly become worse? Think about it. Everything written before 2017 was made by humans. After 2017, LLMs started filling the open web with their own content. And now that same content is being fed back to train the next LLMs. An LLM can give you perfect grammar and a rich vocabulary, but it can quietly lose the real context. Right now the share of AI content is small. But what happens when it keeps growing?
Have the last-generation LLM pre-screen new training data for the next generation LLM.
Aka, due to their mistake. When Sony originally signed an agreement, they should have insisted on a perpetual license for anything already in the customer's library.
I was initially inclined toward some minimal sympathy for Sony here, but I see no good faith reason why they'd sign a licensing agreement which allows the other party to do this.
I used to work in web development and I still do some programming as a hobby, but I'm currently a fifth grade teacher (because I love working with kids).
My email is my username and gmail. As long as you're not a bot and you're not trying to sell me something, I'd really love to hear from you!