gok

It's a fascinating design but it's been 14 years since the concept was first announced and it's never really completely worked. If it ever was possible, it's not clear the talent for it still works for SpaceX.

Ok that's...just cheating. You can't take a benchmark like MMLU designed to test the performance of a single general language model and compare it to performance of a small specialized model designed to do well on MMLU.

An HTML5 super jet?

consider using fp16 or bf16 for the matrix math (in SME you can use svmopa_za16_f16_m or svmopa_za16_bf16_m)

Do you use fully bidirectional attention or is it at all causal?

The oldest iBook G4 is from October 2003, not even 23 years old.

There are many marginally employable swing voters who vote Republican when they have jobs and the Democrats in charge ask for taxes, then vote Democratic when they get put out of work and need a social safety net.

it is possible to put 500 to 1000 TW/year of AI satellites into deep space, meaningfully ascend the Kardashev scale and harness a non-trivial percentage of the Sun’s power

We currently make around 1 TW of photovoltaic cells per year, globally. The proposal here is to launch that much to space every 9 hours, complete with attached computers, continuously, from the moon.

edit: Also, this would capture a very trivial percentage of the Sun's power. A few trillionths per year.

Yes.

While they provide data privacy, our research shows their weaker reasoning and alignment capabilities make them easier targets for sabotage.

If you are using any LLM's reasoning ability as a security boundary, something is deeply, deeply wrong.

The real lesson here should be that doing crazy shit like swizzling the program counter in a signal handler and writing your own assembler is not a good idea.

tldr it's somewhat comparable to an A100, which was released in May 2020.

This has been done in fact: https://doi.org/10.1016/j.sysarc.2024.103102

How hard would it be to add generational collection?

This is like asking a human how many pixels appears in the word "blueberry".

So this 4B dense model gets very similar performance to the 30B MoE variant with 7.5x smaller footprint.

Java is not memory-safe in the Rust sense.

It's emissions credits, saved you a click.

CTA stopped having conductors over 30 years ago. Peak ridership and performance was in 2012, decades after the move to OPTO.

It's kind of too bad Linux to doesn't just support multiple base page sizes.

MoE is in general kind of a stupid optimization. It seems to require around 5x more total parameters for the same modeling power as a dense model in exchange for around 2x less memory bandwidth needs.

The primary win of MoE models seems to be that you can list an enormous parameter count in your marketing materials.

Curious how this compares with, say, the implementation of gemm_s8s8s32 in Intel's MKL / OneAPI.

Which mean Tencent (Epic's parent company) will finally be free to open their app stores world-wide.

Consider adopting `os_sync_wait_on_address()` on Darwin for your futex needs

previously https://news.ycombinator.com/item?id=39485940

I would have assume the major benefit to precision is that it enables compaction…

Computer security is not a serious field. There is no other group that honestly feels "do what I meant, not what I said" is a sign of someone else's bug.

It's hundreds of milliseconds to do a a million iterations. A single time check is hundreds of nanoseconds.

So these guys are just dumping confidential tax documents onto OpenAI's servers huh.

see also https://news.ycombinator.com/item?id=39485940