zarzavat

What's the largest semiprime it can factor?

As the recent "proof" of the Collatz conjecture shows, that's not enough in an adversarial context. Human mathematicians don't submit proofs that take advantage of soundness bugs in Lean. AIs do.

Are you talking about Japan? Because when I think of "Asia" strong job protections are just about at the bottom of the list.

Humans are evolved to survive in the wild. We are not evolved for circuit design. Yet we can design circuits because evolution found it easier to develop a general problem solving nervous system than a nervous system which is adapted for every single specific problem a human might encounter.

> Humanity's Last Exam (w/ tools)

This is one of the only benchmarks that actually matters for testing the frontier however. Other benchmarks can be gamed by simply being more persistent, but HLE is a diverse set of open-ended research-level questions. It tests domain knowledge and problem solving skills. Burning more reasoning tokens may help somewhat but not as much as e.g. coding benchmarks.

Therefore: the current transformer architecture is fundamentally incapable of AGI because the models have no mutable long-term memory.

You only have weights (large immutable memory), or context (small mutable memory).

Humans have mutable long-term memory: I can learn a new skill, adapt an old skill to new information, or learn new knowledge today that I couldn't perform/didn't know yesterday. I don't have a training cutoff.

Context engineering is an attempt to paper over this limitation. You can get really far with context engineering and huge models, but you will never get to AGI because there are many tasks where humans' mutable long-term memory outperforms.

For example, a human can invent a new musical instrument and then learn how to play the instrument they just invented. That's inference (inventing an instrument) leading to training (neuroplasticity). Humans have the ability to train our NNs with considerably fewer training samples. Everything that you can do with transformers is in one causal direction: training -> inference.

You jest but generating SVGs requires understanding of color, size, placement. It's stress testing visual/spatial/artistic capabilities that would be required for writing CSS/design work.

Yes if you're doing backend the pelicans are probably completely irrelevant but if developing anything with a UI, you probably want a model that understands the relationship between code and what the user is seeing.

You can see immediately that it's vibecoded so they kind of don't have a choice but to say so. It is interesting to see what can be achieved with AI even if few people want to play 100% vibecoded games.

I'm polite to LLMs. It's not for the models it's for myself. If I start being rude to models then I might accidentally start being rude to other people as well.

> This time it's a reCAPTCHA successor, which requires the user to scan a QR code using a Google/Apple owned phone.

I can't wait for the future where I need a burner phone and a robotic arm just to browse the internet.

Pizza dough is usually not rolled, it's stretched. If you roll it then you will press out the air bubbles which makes the final crust denser. That's not always a bad thing and some pizza makers do swear by rolling, but it does mean you end up with a different pizza than an average human would make.

It's easy to automate making shitty pizzas (Pizza Hut, etc), but real pizzas require quite some dexterity and it seems it would be difficult to automate. Perhaps start with burgers, that would be easier.

Chess is a brute force search problem. Humans are not good at chess, even a small computer can beat Magnus Carlsen.

It would be better to compare models at how well they can write the code for chess engines, otherwise it's just saying that Fable is not a good CPU emulator, which is obvious.

The objective of RL changed circa 2024.

The 3.5-4o era models were trained by RLHF primarily to write in a way that's pleasing to humans.

Starting with o1 the focus switched to reasoning, coding and benchmarks. If you remember when GPT-5 came out it introduced a new "colder" writing style (which was partly intentional) which made a certain subsection of Reddit Very Upset that Sam Altman had killed their AI boyfriends and they forced him to reinstate the warmer 4o.

The artificial default voice of modern models is optimized for reasoning efficiency rather than understandability. Unfortunately the artificialness has a tendency to leak out into the output tokens because reasoning tokens and output tokens are all still just tokens. You can prompt recent models to write in a more understandable style and they do a fairly good job of it, but it's not the default anymore, which means you will have to keep re-prompting it.

I don't think it's an unfixable problem, Anthropic in particular are just not very motivated to fix it because Claude is much less popular for chat than ChatGPT is.

If you spend all day talking to Claude then its mannerisms begin to grate on you like fingernails on a chalkboard.

There's nothing wrong with the English per se, it's grammatically correct, it just sounds so "Claude". Claude talks using pithy word soups that lack the niceties and structure of human writing.

I thought 'how bad can it be?', yet only three paragraphs in "a specifier now names a repository rather than choosing a transport"

If you're going to serve us slop then at least serve us good slop, give it a pass through GPT 5.6 to clean up the English.

Even Artificial Analysis has Opus 5 better than Fable in their aggregated "Intelligence Index" which combines 9 benchmarks. Opus 5 is heavily benchmaxxed.

It's China. It's a given that they use your data for training. At least they're nice enough to be honest about it.

3. Deployment problems unrelated to the weights causing degraded performance

Freudian slip.

The critical difference for me is if they ask while you're eating, perhaps 5 minutes after serving your table, that's OK it means they are genuinely interested in improving your experience and it's an opportunity to correct minor problems.

If they ask after you've eaten, that would be rude, from my British perspective.

I could say the same about US restaurants. I want to socialise with the people I'm sharing the meal with, not some random waiter. The servers are constantly interrupting, ruining the mood and you feel unwelcome like they want to get you out the door the moment your plate is empty (because that's exactly what they want). It's like being on a conveyor belt.

It does it quite consistently on higher reasoning.

I believe they have made display answer a tool call so that it can keep updating its answer as it thinks, instead of doing all the reasoning up front, essentially inverting the usual flow. You can use this pattern with any model that supports tool calling.

Even if it were possible it wouldn't change the outcome. China is capable of training frontier models even without distillation. Distillation is only an accelerant.

The primary resource you need to train LLMs is money and China has plenty of that.

It is possible to ignore laws if it's too much bother to comply with them... it's hard to imagine the authorities will come after someone for mailing out 10 anti-static bags.

ChatGPT does this. It will present an initial response, then change its mind partway through the thinking process and present a different response.

I'm not saying that Anthropic and OpenAI are definitely dishonest, there is no proof of that. However, when you have

1) an extremely high financial incentive to be dishonest (trillions of dollars),

2) low-ish chance of being caught, especially if you launder the data through another model to remove identifying information,

3) the people in charge of said operations are generally agreed to be snakes,

then it should at least arouse suspicion. You may be right that get enough data from opt-ins that they don't need to do it. I do believe that they don't violate enterprise ZDR agreements, but for normal subscriptions I'm much less confident.

Yes your Chinese open model could have a time-release backdoor, just as your Chinese vibrator could have a hidden microphone that records everything you say and transmits it to the CCP. But does it? No.

What's much more likely is that your US AI provider is promising not to train on your data but is doing so anyway. With a self-hosted model you can at least avoid that.

You have to keep your session warm in cache. Keep the AI talking/thinking. If you have not touched a session for few minutes then /clear and start a new session.

Providers will generally keep your session in cache for at least 5 minutes, possibly hours. The exact cache policy depends on the provider.

If your session expires from cache then the next time you send a message you will have to pay for all the tokens you had used in context up until that point again. e.g. if you have 200k tokens in context then if your session goes cold and you send a message after expiry you will have to pay for those 200k tokens again.

With 1M contexts especially you have to be extremely careful that you don't end up resubmitting requests for hundreds of thousands of tokens again and again.

Try to get yourself and the model to use disk for medium-term context rather than model context, that way it's much easier to /clear and restart if you need to go to the bathroom or something.

You may be causing a lot of cache misses. You have to use the caching efficiently otherwise you can burn up any plan in any amount of time.