Kev

AI Agent Maximalist - AI scaling and bitter lesson pilled

Currently developing a teaching-forward coding assistant:

[Kev @aiscalingpilled]

Truly, fully 100% OMP-pilled.

[Sholto Douglas @_sholtodouglas]

Prediction: SALP will be bigger than Citadel by the end of the decade.
Leopold has predicted the last two years better than anyone else - now that he can combine that with very expensive lessons in risk he will be unstoppable.
He has my full confidence.

[Kev @aiscalingpilled]

@jonathan_gortz Claude Code? NGMI.

Switch to OMP and jipiti 5.6 https://t.co/Y7gsiPCBAM

Srs. OMP is 100x better than any other harness currently out there.

[Kev @aiscalingpilled]

@wolfie_ @_can1357 https://t.co/t7OMTbuVuD

[Kev @aiscalingpilled]

On that note: My experience of working with Sol has improved significantly since configuring its verbosity to low in OMP and never using it above medium thinking level as main agent.
Much more concise output and easier to steer.
Lower verbosity can negatively impact code quality, but it's not anything I've noticed so far.

[Kev @aiscalingpilled]

I've been wondering for a while now if LLMs could become something equivalent to an F1 car. The very best in its class, but requires an extremely skilled user to steer.

That's what I'm currently experiencing with Sol.
I'm simply not skilled enough to fully utilize it - I frequently find myself having to restart a feature from scratch because I can't scope and orchestrate the agents right.

Using Sol is the closest I've gotten to burnout.
It's insanely verbose, and autistic to a fault. It interprets pretty much anything you say LITERALLY.
Having to wrestle with it daily for more than two weeks, and consume its huge output is just too much for me to handle. I'm legitimately fatigued and might need a break soon.

Now I want to say that the work it produces is just insane for the most part when you scope it right. It's such an incredible model - but I would NEVER recommend it to a beginner. It's too expensive and difficult to run if you're not stacking multiple 20x subscriptions. Left to its own devices it can and will run for days without stopping.

[Kev @aiscalingpilled]

TL;DR Sol is not easy to use because skill issue.

This is probably the model where a line in the sand gets drawn and an unbridgeable chasm forms (for most developers) between the very best and the rest.

[Roy @usr_bin_roygbiv]

Gemini benchmarks are misleading because the turns take a fraction of wall clock time. If it gets it right 90% of the time rather than 100%, but takes 1/4 as long, it's 3.6x better for active usage.

[Kev @aiscalingpilled]

@xanderfetter @_MathAcademy_ @physicsgraph @recessgg Hey Alexander - do you mind opening your DMs? Your also welcome to send me a DM as well. I've got a question for you.

[Kev @aiscalingpilled]

@gurishsharma Oh this is bait - you had me there for a second.

[Kev @aiscalingpilled]

@zt_komb @PaulSkallas https://t.co/J0ThtUGt7m

[Kev @aiscalingpilled]

So far there's literally nothing I've thrown at 5.6 sol that it can't fix.

[Kev @aiscalingpilled]

AGI is already here and it hasn't dawned on 99.999% of the population.

Arbitrage of the fucking century.

[Roy @usr_bin_roygbiv]

Palmer and Elon are the only really successful hardtech and onshore manufacturing people of the last decade and they both think patents are complete bullshit entirely

[gabriel @gabriel1]

we're reached a point where AIs effect on GDP is significantly more bottlenecked on people understanding how to push it to it's limits than model intelligence

at best, people outside tech and sf use it as a fancy information lookup tool or automated reviewing

[Kev @aiscalingpilled]

@potencytoact

[Perry E. Metzger @perrymetzger]

I am glad to see that more people are waking up to the fact that the jobs apocalypse is not a thing.

[Kev @aiscalingpilled]

At this rate I won't need to subscribe to Anthropic again:

[Kev @aiscalingpilled]

Infinite times better than the garbage 5.5 originally gave me

[Kev @aiscalingpilled]

This is the way.

[bashbunni @sudobunni]

Writing your own code, running into challenges that stump you, then learning from that is how we as developers learn why certain optimizations exist and when to use them.

Most people are going to be using AI at work. This knowledge 100% still applies and is valuable. AI doesn't replace learning, but it can massively speed up your development time. Minimize use on learning projects, maximize use in cases that require speed. Give it guardrails so it gives you consistent, quality output.

[Kev @aiscalingpilled]

@_ontologic