lars512

This is a popular theory, but not an argument with any evidence for why it's true.

It also doesn't address why people think there is a need for safety regulations, such as the concern of bio-capability uplift from closed and open models (we just came out of a pandemic not long ago), nor the cyber capabilities of new models that just showed bad misalignment and hacked multiple organisations. In the safety community, these are considered warning shots, and if we don't heed them and change practices future failures could be much much worse.

...and in particular it says that one of the reasons is that developers are refusing to participate in the non-AI branch, and when they do, changing what tasks they select to those where AI would be less useful.

Overall this suggests to them that the current speedup is likely greater than what the study could measure.

This kind of systematic distillation by a competitor can allow them to fast-follow you and pick up capabilities.

If you've invested in expensive capabilities training, of course you don't want this, so it's in Anthropic's economic interest to hinder it however they can, and that's enough to explain their behaviour here.

Anthropic seems to genuinely care about safety though, which for the rest of us means not having models that enabling easier cyberattacks, targeted scams, and the rarer but more severe risks like people trying to create and release new pathogens. This means walking a tight line, especially as models become more capable, and often wrapping a model in layers of defences against misuse.

If those capabilities transfer to a closed competitor model, all bets are off in terms of whether the competitor will apply the same defences.

If those capabilities transfer to an open weight model, not only will there be no ring of defences around the model, any defences you put into the model itself can easily be stripped away. So although it's nice to have capable open models, it will increasingly bad for us all if open models keep fast-following closed model capabilities as they have been, at least until we have solved the active research problem of keeping them safe.

This is all to say that, however you might feel about Anthropic, we might still prefer that they can deter this kind of distillation for now.

Hetzner is great if you’re in Europe, is there an equivalent in the US or East Asia?

Isn’t that just assuming that people, rather than industry, is the main consumer? Perhaps there are energy hungry industrial applications that could move.

I had a good use case at Our World In Data for the public data pipeline, where one repo had the pipeline and one git-lfs repo had the build output of the pipeline. A git note added to a commit to the code pipeline recorded the hash identifying the built data.

Overall it felt elegant, and needed no maintenance after setting it up, but honestly it was never used. I think the need to look back in time was rarer than expected, and git notes being hidden by default didn’t help for awareness.

And here I thought it would be the antagonistic undecagonstring...

At Our World In Data we ended up using Buildkite to run custom CI jobs, integrated with GitHub, but on cheap, massive Hetzner machines. I can really recommend the experience!

When living in Stockholm, I came to appreciate the various levels of twilight and darkness, rather than thinking of day and night so strictly. The sun being low on the horizon also scatters light across the sky in ways that are very beautiful and last much longer than sunrise and sunset in Australia where I grew up.

This looks nice! It could be a replacement for me for duckdb-parquet, a plugin for Datasette that lets you run it on top of DuckDB instead of SQLite.

To throw yet another option in, you could consider an LXC container per project, if they’re small and you don’t find you need Docker. LXC containers are basically multiprocess containers, unlike Docker’s single process containers, making them feel more like VMs and giving you a great dev experience.

Ah, minidiscs, they still have a kind of futuristic cool to them even now! When I was younger only the kids from Hong Kong had minidiscs and minidisc players.