benob

Is there anything preventing them from using heterogeneous memory chips, like 1/4 GDDR7 and 3/4 LPDDR? It could enable new MEO-like architectures with finer-grained performance tuning for long contexts.

This work is a good argument against memorization of information seen less than 250 times during training.

I wonder if one could store only the binary representation at training and sample a floating point representation (both weights and gradient) during backprop.

There was a time when people would estimate n-gram probabilities with feed-forward neural networks [1,2]. We just improved that with the (multilayer) attention mechanism which allows for better factoring over individual tokens. It also allowed for much larger n.

[1] https://jmlr.org/papers/volume3/bengio03a/bengio03a.pdf

[2] https://www.sciencedirect.com/science/article/abs/pii/S08852...

Like it or not, LLMs are effectively high-order Markov chains

These techniques are limited to structures that can be checked with bounded history or bounded memory (that can be checked with a grammar or FSA). What about more complex structures that don't factor easily?

My understanding was that the Alpaca data was a distillation from text-davinci-003

a good benchmark for video understanding in IA

It's funny because automation is the only thing you can expect from AI

> YouTube views seem to have fallen off a cliff recently

So they started discounting AI data collection bots?

Stop thinking about text being the data. There are so many other sources, even some that you can generate.

https://arxiv.org/pdf/2506.20057

Even if it's not on topic, that post is quite interesting.

How is it different from mere syntactic sugar over the same programming concepts? What does it bring that C++ cannot do?

Isn't it just a way of controlling the language vs using normative bodies?

Are there efforts to include the neccessary context in compilers to autovectorize?

I wonder how difficult it would be to bias a model so that it subtly corrupts election results when performing OCR.

You can generate 1.0 programs with 3.0 programs. But can you generate 2.0 programs the same way?

Cray1 should be compared to nowadays raspberry pi pico 2 / rp2350 which has similar specs (using external ram).

> This is not an advertisment for Claude Code. It's just the agent I use at the moment. What else is there? Alternatives that are similar in their user experiences are OpenCode, goose, Codex and many others. There is also Devin and Cursor's background agents but they work a bit different in that they run in the cloud.

What do you recommand to get a Claude-code-like experience in the open-source + local llm ecosystem?

Watermarking is easily disabled in the code. I a wondering when they will release model weights with embedded watermarking.

The gemini web (smolweb) has no effective search engine, and therefore links also play a crutial role in content discovery...

I am seeing different prices everywhere: - $199 https://shop.vassarrobotics.com/ - $219 https://shop.vassarrobotics.com/products/navrim-robot-that-l... - $599 https://www.youtube.com/watch?v=TbDTCwzFeIU

Which one is the actual price?

Isn't it time for disallowing browsers to connect to anything outside same origin pages except for actual navigation?

Servers can do all the hard work of gathering content from here and there.

It looks pretty and featureful, but I'd love a cheaper version, maybe without a display.

I wonder how to import markdown into Notes

I guess autoregressive llms can be finetuned (or continual-pretrained) to do inference using diffusion. We've seen a recent paper (which I don't remember) training from scratch, but it seems overkill. Do Google say how they did it?

Also, does diffusion have the potential to increase speed of cpu-only inference?

Ok, one of my email addresses is in a bunch of leaks. What is interesting is that most services on this list I have never used. How did they get my email in the first place? What is the accuracy of that whole business?

It looks like it is an extension of 3d model generation techniques.

I wonder whether they used a mirror or simple recordings as baseline

> If you ever remove the interconnection PCB, make sure to put it back with the same orientation. It will fit just fine when rotated 180 degrees but the scope won’t work anymore!

I remember pulling a 486 out of its socket in the 1990s and putting it back with the wrong orientation. There was a poc and a bit of smoke. Something on the mainboard had burnt and it wasn't working anymore.

I used smell to locate the fault, a big trace on the PCB, which I soldered back and magic, it all worked again...

I think you are mixing things. Quaternions are in the same category as complex numbers. They can be represented as matrices, and there are probably nice uses of matrices where the element is a quaternion (such as QDNNs) instead of a real number. My experience is that in massive architectures such as LLMs, simpler forms are more successful unless there is a true benefit to representing things with more elaborate types of scalars (such as in physics, or 3d graphics).