I’ve run some underwhelming local LLMs and done a bit of playing with the commercial offerings.

I agree with this post. My experiments are on hold, though I’m curious to just have a poke around DeepSeek’s stuff just to get an idea of how it behaves.

I am most concerned with next generation devices that come with this stuff built in. There’s a reactionary sinophobe on youtube who produced a video with some pretty interesting talking points that, since the goal is to have these “AI assistants” basically observe everything you do with your device (and are blackboxes that rely on cloud hosted infrastructure) that this effectively negates E2E encryption. I am convinced by these arguments and in that respect the future looks particularly bleak. Having a wrongthink censor that can read all your inputs before you’ve even sent them and can flag you for closer surveillance and logging, combined with the three letter agencies really “chilling out” about eg Apple’s refusal to assist in decrypting iPhones, it all looks quite fucked.

There are obviously some use cases where LLMs are sort of unobjectionable, but even then, as OP points out, we often ignore the way our tools shape our minds. People using them as surrogates for human interaction etc are a particularly sad case.

Even if you accept the (flawed) premise that these machines contain a spark of consciousness, what does it say about us that we would spin up one-time single use minds to exploit for a labor task and then terminate them? I don’t have a solid analysis but it smells bad to me.

Also China’s efforts effectively represent a more industrial scale iteration of what the independent hacker and opensource communities have been doing anyway- proving that the moat doesn’t really exist and that continuing to try and use brute force (scale) to make these tools “better” is inefficient and tunnel visioned.

Between this and the links shared with me recently about China’s space efforts, I am simply left disappointed that we remain in competition and opposition to more than half of the world when cooperation could have saved us a lot of time, energy, water, etc. It’s sad and a shame.

I cherish my coding ability. I don’t mind playing with an LLM to generate some boilerplate to have a look at, but the idea that people who cannot even assess the function of the code that is generated are putting this stuff into production is really sad. We haven’t exactly solved the halting problem yet have we? There’s no real way for these machines to accurately assess code to determine that it does the task it is intended to do without side effects or corner cases that fail. These are NP-hard problems and we continue to ignore that fact.

The hype driving this is clear startup bro slick talk grifting shit. Yes it’s impressive that we can build these things but they are being misapplied and deferred to as authorities on topics by people who consider themselves to be otherwise Very Smart People. It’s… in a word… pathetic.