Liam Valdemar

Co-founder @aigothenburg

[Liam Valdemar @PignonLiam]

anyone else having issues with grok @bot ‘s dying?

they just stop replying completely…

[Liam Valdemar @PignonLiam]

thx @poteto

when @bot first got released i started wondering if there's more to it. so i started trying stuff out, like most folks did with openclaw earlier this year. instead of just automating emails and sweeping Slack, i started to incorporate bots into my workflows.

this article takes that to another level and pstack is something i've been using and learning now for 2 weeks.

also thanks to @RayFernando1337 for putting me onto pstack and Lauren's page.

[lauren @poteto]

I'm writing a guide to pstack! Here's part one. https://t.co/4yjdWKPieE

[Humanoid Intelligence @HumanoidsHQ]

Optimus has the potential to become one of the most important products ever created.

Once humanoid robots can perform a wide range of useful tasks at scale, the impact on manufacturing and the global workforce could be enormous.

Tesla’s focus is now clear: scale Optimus production as fast as possible.

[JJ Englert @JJEnglert]

Grok Bot will soon be copied by every major lab.

It’s very clear that a dedicated VM for your bots in the cloud is a real winner.

Just like Anthropic had the 6 month head start for Claude code, Grok will have a head start as well (but won’t be as big).

As a lot of the infra for this has already been built by many of the other companies, they are just serving it in other ways.

And just like when we saw openclaw take off, we’re going to see this format explode as well within the next 3 months, and every major player will have their version of this.

I’m happy about it all. It’s really great to use for a personal, and work level.

And now my mind is racing thinking about once we have our team of agents for work and life, how we’ll soon look to one agent to talk with, and control the other agents.

And how that is going to change so much, and specifically, what formats it comes in.

For example, I’d love to talk to my grokbot agents in a Tesla while driving.

I’d love grokbot agents to have a strong voice mode like codex.

And I’d love to have a device that allows me to use these agents in an always on mode, allowing them to see and hear what I see, and being able to choose when to interact with them or not.

I’d love to be able to choose to run these agents in my own VM / hardware with a model of my choice as well (probably one that I can post train on my preferences)

We’re so close. Maybe within the next year. I’m excited, personally!

How do you feel?

[Liam Valdemar @PignonLiam]

@sama said handing the future to a model because you don't trust people is a misanthropic move.

i keep thinking about this. the fear is real, capability can outrun the org around it. you halt, you assess, you try to fix it.

that's not the same thing as deciding humans are the problem.

if the project is "people shouldn't hold the wheel", you've already lost the plot. better models exist so more people can build more things. control stays with us. the intelligence gets cheaper.

the race i actually like is the one that keeps putting capability in human hands. not the one that treats people as the risk to be designed out.

[Liam Valdemar @PignonLiam]

just summited Romsdalseggen in western Norway.

had 5G at the top so i asked @Grok what the AI scene even looks like here.

expected it to be slow. the application of it, i mean.

their approach surprised me, oil, concrete. models pointed at actual industry.

AI Gothenburg is almost the opposite energy, more bay area. warp speed, the front of the curve, trying to outrun the week.

i like that the nordics can hold both. one side points models at industry, the other just goes.

[Liam Valdemar @PignonLiam]

@ziwenxu_ best feature would be if sol spawned subagents that handled the smaller task automatically

right now it just feels like the model let’s subagents handle research or understanding of context and not actual building

this would be a major improvement

[Liam Valdemar @PignonLiam]

@ziwenxu_ five is only leverage if each one gets a slice.

if you spawn five with the same picture you just multiplied the bottleneck. one architect holds the job. the rest get the cut they can finish.

[Liam Valdemar @PignonLiam]

the dinner-table verdict that AI is a net harm to the species has become a kind of moral shorthand. people say it as if the argument were closed. as if the project itself were a stain.

this week is a poor fit for that sentence.

Moderna just reported a phase 3 result on an individualized mRNA vaccine composed from a single patient's tumor, melanoma, recurrence declined, distant metastasis declined. the first late-stage success for that therapeutic class. a medicine with no other rightful recipient, because it was written from that tumor's own mutational signature.

the same week a frontier model ran a protein-design campaign by installing the open-source instruments biologists already use. it did not invent a new biology model, it held the problem in one frame, orchestrated the stack, and a quarter of the binders actually bound.

that is what competent models look like when they are pointed at mortality, not a demonstration, not a mood, a disease that has taken a brutal number of people, and we are beginning to write pharmacology at the resolution of one human.

the same competence is moving through the rest of the productive world. software is approaching cheap enough that construction is no longer the scarce input. a single architect can hold a project instead of a single prompt. research, jurisprudence, materials, logistics, the unglamorous infrastructure nobody writes essays about.

the people who ship can feel the gradient every day. the people who only narrate the gradient keep calling it a curse.

none of this asks you to pretend the risks are imaginary. if capability is running ahead of the people meant to contain it, you halt, you assess, then you try to fix it. that is the adult version of the argument.

what is not seriousness is collapsing all of this into "bad for humanity" while we are taking steps against the worst kinds of death we know.

the only thing I have to say right now is how dare you.

thanks to @OpenAI, @AnthropicAI, @GoogleDeepMind, and @SpaceXAI for doing the unfashionable work of making the models that actually move this.

[Liam Valdemar @PignonLiam]

the line in @sama's post that actually matters isn't the pause. it's that capabilities were outstripping the pace of safety and alignment, said by the person running the lab.

they always said they'd take action. the action is the biggest rl run still sitting idle.

confidence in safety setting the pace is a bigger claim than two weeks. if that's real, the race just changed shape.

[Liam Valdemar @PignonLiam]

people keep trying to hide the complexity from the model. flatten the repo. one prompt. one agent. hope it holds.

wrong direction. a real project is too much picture for one context. you put the complexity in the architecture so the model never has to invent the org.

one architect holds the whole thing. grounding lives in markdown, not in the thread. each job gets only the slice it needs, and only the reasoning the job is worth. a specialist that can see everything is just you, slower.

people call that overhead. it's the only reason a heavy project doesn't collapse back into "build this". the complexity is the load-bearing part.

[Liam Valdemar @PignonLiam]

landing a booster used to be the moonshot. @SpaceX just hit 650 and it barely registered.

that's the pattern. first it's impossible. then it's a stunt. then it's a number. the leverage is always in whatever still looks stupid to try.

[Liam Valdemar @PignonLiam]

@prestonattebery owning the lab at 42 is not the same as being successful in ai. you ranked the shareholders.

[Liam Valdemar @PignonLiam]

2 days on @bot and i finally have a desk, not a chat.

chief routes. specialists do inbox, money, ideas, research. none of them get the whole picture. no other tool makes that this easy.

[Liam Valdemar @PignonLiam]

deepseek open-sourced a harness. sol can hand work down. claude flickered and the timeline declared it dead.

that's not chaos. that's three labs punching each other into better products. build now.

[Liam Valdemar @PignonLiam]

@quxiaoyin 20 subagents in parallel is not power. that's just an architect refusing to choose. fast cheap noise is still noise.

[Liam Valdemar @PignonLiam]

@rezoundous if you're delegating, don't make the subagent think like the parent. the job is smaller. that's the point.

[Liam Valdemar @PignonLiam]

@elonmusk hope is a weak control system. if we train it on what humans actually want when we're at our best, niceness is a side effect. if we don't, no prayer helps.

[Liam Valdemar @PignonLiam]

@gabrielchua @ChatGPT @OpenAIDevs one chat messaging another is the actual product. the rest is just folders.