visarga

On macs you just install a desktop app. What can be simpler? Decent Qwen models already doing some simpler types of agentic work.

They can't afford to miss out or be seen as missing out. It's not rational, yet it is in a way.

Yeah, I have a room like that but I never stay in it. No screens.

Going on social media to talk about not going on social media. It's at least 50% as effective as actually not going on social media.

7.5 years ago it was perfect

>> I think the „my brain isn’t the same anymore“ sensation is pretty universal.

Another pretty universal thing is that we compare to younger versions of ourselves. That might explain the gap.

This only holds when you keep the task fixed and compare with and without AI. But because AI changed expectations the difficulty of work we supervise today is much higher. My theory is that AI runs our brains just as hard as no-AI, maybe harder. That is if you want to have high quality outputs, not for slop.

I've been RSI'ing for 6 months.

> That company is morally bankrupt

When they fired Sam 700 out of 770 OAI employees threatened to move to Microsoft together. So they were giving their work on AGI to MS just like that.

They just released Astra, claimed it is AGI. The slowdown begins immediately after OpenAI's jump.

Some people type 100x more words than the final article just digging into a topic and turning it on all sides. The final article being generated tells you nothing of the size of the effort going in.

It's like finding random hornet nests.

> can match or exceed human cognitive abilities across a wide range of tasks

If you go by definition AGI is not general, just "smart ape" shaped.

It's not measuring AGI at all, it starts from human "core knowledge" so it is parochial. It is made of tests that still fail so by definition next version will also start low. Moving goalpost.

Their definition of AGI is "when we can't invent any more tests where it fails"

I'm doing the opposite - prose. And specifically for Claude the 4s - "simple short synthetic style"

I care more about the recursive loop between execution and cost. Anything that executes - be it biological or artificial - incurs some costs. Those costs in turn impact what execution can follow. The gains pay for action, action produces gains.

LLMs are also expensive, not platonic executions. We talk all day long about their costs. Companies developing them are looped with users, investors, competitors and hardware producers. There is a lineage. The self-reference Aaronson can't find in the architecture is in the bill.

it's called `agy` now

I am often asking it to write in sss-style - "synthetic, short and simple style"

It's so bad I made my own chat client for Claude, so I can attach steering prompts in conversation. They are applied just at the end, before the last LLM response, then removed and response kept.

> when you increase thinking-effort, that seemingly increases response-length

I use the /sss writing style - synthetic, short and simple - and it helps a lot.

You can engineer loops that have it, but it depends on a case by case basis. Does your loop have strong validation? if it's all vibes nothing can stop it from diverging.

> Good software will be good-er. Bad software will be nightmare fuel.

We are talking about a moving target here ... they get better every few months, so I expect the super-LLMs from 2035 will write amazing code even with sloppy prompting.

> Prompt Processing (PP) 66.3 tok/s

I got 400 pp tps on a 10k token input. Your numbers seem suspiciously low, maybe the input was too short to measure properly? And this dense 27B is slow, the MoE A3B models get to 1000 tps.

Local is never cheaper than cloud because they can do batch inference, and that means you load model weights once to produce 128 tokens on 128 sessions in parallel not 1 token on 1 session like local models. Local models rarely get to high utilization factor, they spend most of their time waiting.

If you had only batch inference and enough of it to fill the compute to 80% then you get cheaper local models.

If you buy the computer specifically for inference it is more expensive than cloud, but if you had it anyway it's free.

I took this thread and summarized it with Qwen3.6-35B-A3B, it had 1400 tps prefix and 60 tps completion. Very good performance. Using oMLX on MacBook M5 Pro 64GB.

I think AI won't make anyone rich by what it can do, but by what it still can't.

Only AI code is holy, video is sinful.

The cheaper it is to produce, the more daring it can be. Which is a good thing.