fancyfredbot

The product makes the money which pays for the ads which pay for the attention.

I tried to reproduce this with an independent test. I find Claude can interpret small networks reliably. However it can't reliably propose improvements to even small networks.

I had Claude train three language classifiers for short words in four languages. The network was just over 10k weights with a single hidden layer.

I uploaded 6 language word lists and made it train networks to fit random subsets of 4 languages. The languages were picked randomly. Claude could always identify which 4 languages were chosen and in which order. It told me that the language heads were easy to link to letter frequency and each language had distinctive makers. So claim upheld, an LLM can interpret small networks.

I then asked Claude to improve the classification accuracy by proposing an edit to the weights in the above networks. It's proposed improvement reduced classification accuracy significantly. The same thing happened if I gave it a deliberately undertrained network with clear room for improvement.

The task was the same - improve math reasoning capabilities by editing the weights. But yes, I agree that in Qwen it could do this through reducing looping behaviour, but Gemma wasn't prone to looping and it didn't find anything to help, possibly because there wasn't much room for improvement.

For the tiny models with a few thousand weights, it seems plausible Astra can actually understand what the network detects and figure out it's e.g. a triangle detector. This is possibly helped by the existence of many breakout playing networks on the web.

For Qwen, even the smallest versions like 3.5-0.8B it's becoming less plausible that it could be interpreted in a forward pass, even with reasoning. I'm wondering if the result here is recalled from many examples of instruction tuned versions of this model.

A smaller Gemma model doesn't fare as well but unclear whether this is due to fewer users and less weight examples or because it's already very well trained.

If you haven't rebooted yet then you can check how many mtcp handshakes you've received. It takes a lot of them to exploit this. 48 minutes according to the disclosure. So if it's a small number you haven't been hacked. Use "nstat -az | grep -i mpcapable"

If you run caddy it's written in go so probably listening on an mtcp socket meaning you are probably vulnerable. Watch out!

The range of opinions is probably healthier than homogeneity. I think labelling people extremists and pushing them towards the centre would be pretty dystopian.

What is easier to change than people's opinions, is the two party political system which tends to resolve such questions to one of the two extremes based on one side achieving a tiny majority, rather than to the average which is usually a lot more sensible.

That's very true but unfortunately Anthropic have locked in contracts for so much compute that they absolutely need to expand beyond the corporations currently paying them.

Realistically they need to 3x their revenue to pay for their compute. While that's plausible, I'm not convinced that it can all come from developers. Many many developers are already using Claude. I think they need consumer, or some other market too.

I took it as a comment about voters than parties.

It didn't even occur to me that it was saying republicans and democrats are the same. That's just ridiculous ;-)

Edit: To clarify there's good evidence that people overestimate the degree to which people on the other side of the political divide hold extreme views by about a factor of two. We genuinely are less different than we think. It's not necessary for people to be the same for this to be true.

For people with IQ over 80, highlighting that people in different tribes are actually alike doesn't count as social commentary?

Not like the banks. Those are carefully scrutinized by large well funded regulators and auditors and subject to complex legal requirements developed over decades.

To be fair tether hasn't actually lost billions of dollars yet, whereas banks absolutely have. Overall though I don't feel inclined to trust most stable coins. The cryptography protects your virtual currency but the governance is what protects the assets backing them, and the governance looks weak.

Why is it showing me $749 excluding tax? Did the price increase by $50 since this was posted?

Some great writing here. The irony of the guy campaigning to delete a personal page having had his own page deleted is poignant. It's hard not to read something like identification with the aggressor into this.

They didn't pay the Nvidia tax

Makes sense now. Thank you!

I'm sorry if I'm being dense here but I can't see how that would work?

The spec on your GitHub says: "Two simultaneously-live resizable values never share a stack."

So I surely can't have a dynamic array of dynamic arrays share a single stack? That's an unbounded number of resizeable values sharing a stack?

You could make it work on a single stack but that would involve having non-contiguous memory for each inner array since you can add to any array in any order.

The spec implies that we need one stack per element of the outer dynamic array which would let you have contiguous memory for each inner array, but you wouldn't know how many stacks you needed at compile time.

Thanks!

I'm a bit confused about how you could decide the number of stacks at compile time. I can see you could see how many arrays are lexically in scope. However if the code puts one dynamic array inside another dynamic array I can get to an unbounded number of total stacks without ever needing more than two in lexical scope at once. Do you also add stacks at runtime? Like a stack of stacks?

What happens if the memory is freed in a different order to allocation? Does the language disallow that? Or does it get allowed but mean the stack can't shrink?

How are you deciding how many stacks to use?

Sad that that a "buy everything" strategy performs best. It's not a strategy which varies much based on other players behaviour which makes it hard to counter and therefore boring as everyone has to play that way. I guess where there's still room for adaptation is the auctions if the player landing on the space can't afford to buy. The bidding strategy here forces some player interaction and room for different strategies...

Mimo is very good.

I thought it was 6 Sol which was a rebadged Terra? You think 6.1 Sol is too? It does seem to benchmark better than 5.6 Terra? Regardless agree both models have been very bad for openAI's Sol brand. To be fair it's at least cheap compared to other OpenAI models.

Cool! XLA is pretty good + potentially a better compiler infrastructure than inductor which will lead to improved performance on all hardware.

Also I guess this is the final nail in the coffin of Tensorflow.

There's something really cool about this but also something sad.

There was a time when CERN would have been establishing partnerships with US labs. The US is not a CERN member but has provided a lot of funding in the past despite it's observer status. These days funding is getting cut, the US is not as trusted as a partner anymore and a European partner is seen as important. This is sad.

That said I'm happy for Mistral and I'm sure they'll be a great partner for CERN.

The UK spent 100m GBP on a 900m tunnel. Rather than protecting cars from rocks, this one protects trains from bats (or vice versa really) so structural requirements are presumably less stringent. It's weirdly pretty though.

I think you know how to do this. You are head of engineering at an AI startup. You know how to create and revoke API keys and how to share them with your interview candidates.

At that price I think you'd pick Opus 5.5.

It's perhaps a face saving thing to pretend you'll charge the same. Let's see if it actually happens.

The thing about Zuckerberg is that although he's deeply odd he has no ability to conceal this.

Sam Altman is much more dangerous because he is occasionally able to come across as human.

Of course you need talent but supply seems reasonably high. We have openAI, anthropic, grok, muse, mimo, deepseek, qwen, GLM and Gemini all producing exceptional models.

The biggest differentiating factor seems to be compute. I know most Chinese labs would really like more compute. Google and xAI too, but largely because they sold compute to their competitors.

The point of the knowledge is that it helps you ask the right questions, not so much the ability to answer them.

Knowledge has been available just by asking Google for decades now. The LLM makes it easier but it's a difference of degree not of kind

Until the models are 100% reliable knowledge will be required in order to quickly spot issues and work efficiently with the model to address them.