sarjann

"It is the training apparatus of your future senior staff, and you should protect it the way you’d protect any other piece of critical infrastructure."

I think a problem is that especially with the culture of job hopping a lot and high turnover you can't be sure that you're not just training staff that'll go elsewhere. It's like an industry wide prisoners dilemna.

Why use electricity alone for open models? Surely you’d want to spread the cost of hardware over the period too?

What are you actually trying to get at? Had you even bothered to watch the 2nd (short) yet?

My point was that he’s willing to acknowledge that he was wrong, that is what the clip shows. Also can you read minds now?

‘He's also just saying what he thinks people (the general population) want to hear’

I can't find a counter to every statement they have made and not sure about your specific AGI claim but.

https://www.youtube.com/shorts/QMhtTO3u61A https://www.youtube.com/watch?v=L_ueDUrkOlQ

Examples of acknowledging he was wrong

Not sure about being early when it comes to points about models reaching the peak of their capability. Honestly, I don't understand how he's qualified (in terms of knowledge) to make such claims.

Circular financing is an issue if the wealth accumulation stays in the chip -> model provider ecosystem. However it seems like the labs are making quite a lot of revenue from the chips (customers).

I think one important distinction is at least they have some humility to admit they were wrong. I think Ed Zitron has rarely, if ever acknowledged he was wrong.

I've always been suprised with the amount people take him seriously. I think he's someone who makes people who dislike AI "feel good". I don't blame someone who doesn't understand AI or have read what he's written.

But for the bloombergs and other podcasts I would have expected them to do a bit of research. I honestly think with the amount of doubling down he's been doing that he's a grifter.

I think an important part is communication, so many of these issue could be fixed by saying "Hey we're seeing our utilisation go over x% over the weekdays so we need to implement "surge usage" during this period starting in 2 weeks.

Instead often it feels like they make a change, then wait for someone to figure it out. Then Anthropic ends up being reactive as opposed to proactive in communication.

It is kind of funny because surprises from OpenAI tends to be positive (Tibo resets), on the Anthropc side I dread them.

I guess the Riemann Hypothesis is an easy task then.

That's why it's a choice. If you weren't going to sign up to a site then it's just a pure degradation of experience to see links you can't use.

I would like to be able to whitelist some paywall sites, e.g. I have an FT subscription and would like to have that bypass the block.

Great news, has made low memory bandwidth model usage so much nicer.

Except he tries justifying it.

"Github’s servers are constantly on fire as their usage increased something like 50x due to LLM sloppers pushing large amounts of trash code"

"Unless you have a emergency hotfix (you don’t), go hit the gym or walk outside." with assumptions like above.

If it was, "hey they f*ed up but there's no point having an overly emotional reaction" that would be fine. But he seems to be justifying this. Especially with Github's record up to now of unreliability I think it's completely fair to be annoyed.

This has to be rage bait, this is a critical piece of infrastructure for many people. Stuff going down can lead to deployments failing and as you mentioned in many cases emergency hotfix's.

The idea that we should be fine with this unreliability is just amazing. It's not a mental health issue to have problems when important infra fails.

You would think for a company that has close to a monopoly of developer mind share wrt git that they'd maybe put in the bair minimum to get several 9s.

I think at 1000x it's probably way cheaper to do an "AWS outpost" style thing. Although not sure if it's a fundamental constraint or something that could be optimised.

Introductory price seems a bit weird as it expires at the end of the year and by then it’s going to be significantly outdated.

Might be easier to do cross platform.

Fundamentally I always try to look at incentives. It's not that they are good or bad but that incentives favour certain behaviours.

Google is incentivised to collect a bunch of data (like FB) to improve their ad serving etc. Apple does not have that incentive as they don't make as much off advertising as hardware / app revenue.

NVIDIA is friendly with open source as they want to commoditise the model layer and take the gains in the hardware / data center layer.

I am trying to think what the incentive is here.

If I were to guess, they may have some "test" users (people who are involved in this) who are on the watchlist, they do their regular commutes and see if the system picks them up each time.

For this system there is no requirement that criminals faces are used to test it.

Google + Motorola aquisition vs Apple is a prime example.

I do wonder if we need a charitable foundation that does software patents and makes them freely available just to stop cockroaches abusing patents.

Another step (although governance would need to be super good for this). Would be for this charity to withold access to these patents from companies that decide to abuse stupid patents.

I do think some of features in their harness seem interesting (workers in separate worktrees at once), recovery from crashes seem interesting.

I don't see how not? A theoretical physicist can do all the thinking they want but if they can't test an idea against nature it's not super useful.

dgx spark, nvfp4 so I have spare room for KV cache (context)

Sounds like this would be expensive on the input token side with some saving on the decode side. For workloads with large documents / context that could be an issue.

If it only switches when cache is cold, it seems like it'd lock in one model for the whole session?

What use case are you imagining where cache eviction would happen during it? It might also be the case that a task might seem easy but then it ends up becoming hard for certain parts. For spawning subagents I could see this being useful maybe.

I guess the argument might be that it'll be faster than finding a spot. Slightly minor benefit is you don't need to remember where it's parked.