They’re going to use AI to install the AI.

Recursive Self Installation

IBMs a bit weird. They fired a shitload of them a while back.

Then they said they're only replacing junior roles.

Then I think they're hiring a lot more juniors because they need to build the next seniors.

I didn't really follow too closely but that last statement is at least something I can respect.

[edit] I think mushroommunk is absolutely on point [/edit]

I don't like it, but that would just make them cheaper to hire back as AI babysitters.

Yep. This is why Nvidia is preaching open weights literally every chance they get, too.

Still. It’s a better stance, even if for selfish reasons.

For a company "on-prem" might not mean the same as "local". You might run them in your company data center or in a colocated server (rented space in a data center). In either of those cases, the data center might be physically close to you, but it could be on a different continent.

"On-prem" implies that the company has physical access/control of the server.

"On-premises" as it refers to enterprise hw/sw, has been a phrase much longer than "local LLM"

Even though everyone moved to cloud, most companies keep some portion of their infrastructure on-prem, like prod env in the cloud and on-prem test env/backups, or vice versa

Maybe most consumers wouldn't recognize the phrase, but industry professionals certainly would

It was written by Claude

He should consider making a chewing gum company. Might not last more than 5 seconds though.

If I could pick a company that has fucked up more than Microsoft or EA, it would be IBM. For the past 20 years, the whole company has been on a wild bender of buying good products, and shoving them headfirst into the ground.

It doesn’t literally mean “90% failure rate vs 100%”

In the HLE benchmark, for example, we may be talking about 45% test score for MiMo Pro or GLM or something, vs 55% for Sonnet and 61% for maxed out Claude.

But further complicating that, those particular open weight models do better than Claude in one particular hallucination benchmark, meaning Claude/ChatGPT might be worse at RAG/extraction type workloads. It’s like this across the board, with different models trading blows in different areas.

By “90%,” he really means “it’s so close the better model depends on the particular application, workload, harness, sampling configuration” and such, and that none of these models are close to “100% accuracy” across the board.

One other thing: Anthropic, OpenAI, and Grok have all been accused of “dumbing down” models under high demand to speed up inference. If your application depends on them, and they are the only provider, that could throw a wrench in the workflow.

But open weights? Pick your provider. You get an API as consistent as you are willing to pay for.

In other words, that “100%” isn’t really representing the potential inconsistency one gets being locked to a single provider.

Don’t assume the expensive model is 100% accurate. The cheap model is 90% as accurate as the expensive model, which is already less than 100% accurate.

I thought of it as grading as in an academic setting. If the top commercial models are putting out essays that are graded at 100%. The on prem models are putting out results that are graded at 90%. In the case of support emails, the results won't be much different.