Feed aggregator

We aren't getting to AGI without a fight

Hacker News - 2 hours 42 min ago

I think I realized why I am so angry about the fable takedown. It's because I realize we aren't getting to AGI without a fight. This is the powerful closing ranks to keep capability centralized. This is the ceiling of model capability which we will be allowed to use. Greater capability will exist but it will be gated behind clearances, vetting and power. Reportedly, Amazon relayed these jailbreaks to the administration. Any investment Amazon has in Anthropic is dwarfed by its investment in AWS. The Anthropic investment was likely a hedging of risk and a seat on the board so they could do exactly what they just did.

I've had a legacy migration on the back burner for quite a while. It's a substantial one, where I'm not just translating code into another language with functional equivalence, I'm completely redesigning everything from the ground up. AI has helped a great deal but only in context size chunks. The problem with this, and this is a recurring theme in AI coding is that you need to understand the entire system conceptually and hold it in your head before you can safely modify any single chunk. This fact is the moat between senior devs and AI coding agents (and I don't think it's going away any time soon).

For three days I had access to a model that wasn't perfect but it was making strides in this area. I have been preparing it for well over a year with conceptual documents, explaining the intent behind design choices, constraints that were counterintuitive. Opus and 5.5 Pro do ok tackling this challenge but it was clear that Fable was close to an actual step change in models. I wouldn't assert that it was AGI, but this takedown makes me realize this might mean we've hit the capability ceiling, not because of technical reality but because the gate has been closed by the institutions with the most to lose.

Once the inevitable distillation processes start happening open source will start exceeding the gated frontier models simply because they can't gate open source as easily as open source so that predictable lag of open source behind frontier will erode. But inevitably they will most certainly come for open source as well. I don't think the powers that be or the masses will "win" but like with the printing press we'll land somewhere in the middle. What's clear is that we need to fight actions like this tooth and nail if we ever want to have access to more powerful models.

Comments URL: https://news.ycombinator.com/item?id=48520967

Points: 1

# Comments: 0

Categories: Hacker News

Ask HN: How do you avoid / get out of LLMs local minima?

Hacker News - 2 hours 46 min ago

When using LLMs, I have the impression that they often stick to an idea which they dont seem to be able to shake off, no matter how much arguing I can put in, and however strongly I feel about it.

What's your trick to get the LLM out of its own beliefs ?

I often find myself having to kill the entire context, which is obviously not the best...

Comments URL: https://news.ycombinator.com/item?id=48520927

Points: 1

# Comments: 0

Categories: Hacker News

Red Hat Enterprise Linux Test Page

Hacker News - 2 hours 47 min ago

Article URL: http://supportdownloads.adobe.com

Comments URL: https://news.ycombinator.com/item?id=48520919

Points: 1

# Comments: 1

Categories: Hacker News

Pages