Dave Sumter

@pmarca FOMAI?

We need a name for “I really should be spending all my time talking to AI, but I’m not”. AI pre-psychosis? Or is it just AI psychosis? Asking for a, well, you know…

@RhysSullivan You ain’t lying.. after 25 years making html, I’m still torn between the diff solutions to preserving input state.. all trade-offs..

This is one of the hardest web dev problems and a great interview question

@twlvone Distribution is now infinitely harder than building.. I used think of them as sort of 50/50, now distribution is everything…

@davesumter the constraint isn't building — it's distribution and trust. a one-person company can now build the product. but customer acquisition, legal, finance still need presence. that last 1% you mention is really: does the market trust you at scale?

The solo unicorn isn't happening in 2026... AI closes 99% of the gap, but at unicorn scale, that last 1% explodes into dozens of problems across support, legal, sales, and ops. One person can build more than ever before. But a $1B company solo? Not yet.

Good catch Dave, but I still can’t open the airlock doors….

@tangming2005 Honestly not my experience at all… these days I just post some screenshots, a file reference or two, and ramble into the microphone for a minute.. Claude reproduces it and fixes it every time now. 🤷‍♂️

Single biggest improvement I made to my CLAUDE.md: "When I report a bug, don't start by trying to fix it. Instead, start by writing a test that reproduces the bug. Then, have subagents try to fix the bug and prove it with a passing test."

Running a local model server on my Mac with something like LMStudio + hitting it from ai terminals (Codex/ClaudeCode/OpenCode) occasionally is probably the sweet spot rn. Resource utilization is solid, tech is stable. OpenClaw still too bleeding edge for daily productivity driver 🤷‍♂️ Great hobby project for weekends though. Probably the future..

The level of abstraction you can achieve with AI agents today is just insane. 🤯 I'm completely rebuilding many of my automation workflows to just use an agent with thinking rather than tons of hard-coded rules for doing things. The speed to get complex workflows up and running now is mind-boggling fast...

Just automated the whole house with OpenClaw. Runs 24/7. Supposedly “hands-free”. Wife: “fine” detector (still broken) Son #1: bedtime negotiations like a corporate lawyer Son #2: screen time negotiations like a hostage negotiator Dog: everything is P1 URGENT Neighbour: passive-aggressive updates on the doorbell cam TV: recommends a show, instantly forgets my taste Laundry: washes everything except the one thing I actually needed Dinner: “healthy” optimisation (nobody asked) It planned 12 meals and held a “family stand-up” without me. Then the gateway needed a restart, the local model refused bedtime, and now I’m “Stakeholder” on my own life. 🦞

Holy moly, every day is a new discovery. This is like being on 'computers' back in the 90's... AI is so awesome. Managed to get AI terminals launching directly from my app admin portals today, doing work directly for me.. So simple, why didn't I think of this last year...

@bradmillscan Shit, glad it’s not just me then… I mean, I like it, and I love what it’s going to enable, but fuck it’s sucking a lot of time.. It the moment I’m carrying on with Claude Code as my daily driver and then playing with Claw in the evenings when time permits …

Trying to get OpenClaw agents to do useful work is like trying to win at trading crypto - only the top 1% win. The rest of us end up being the lobster meat for the host in the shell. OpenClaw agents are terrible at executing complex multi step processes that require delegation. I spent about 40 hours last week creating a framework for delegation, standards, accountability, workspace hygiene, commit cadences, capability logs and defining project success. Using all of this I spent 3 hours getting to an extremely well baked “definition of done” for a project to transcribe 200 hours of video content and make it into a semantic searchable knowledgebase. 3 phases, ~10 steps each phase. Clear outcome, clear success cases, clear failure modes, OpenClaw owns the process of doing. Today I am back micromanaging my agent, logged 10 or 15 regressions and generally have frustrated myself into a state of high stress. Constant failure. Fail after fail after fail. 2 steps forward 1 step back. It’s a fucking slog getting any real work done. But the alternative is it just doesn’t get done. I was under an illusion that Agents can autonomously do things for you. Nope. They are a tool. By installing openclaw you are creating a massive workload for yourself. It’s not like how the fantasy threads on here make it seem. I need to level-set my expectations on this tool. 200k context window = it forgets 80% of your processes multiple times a day. Regressions are as common as successes. I’m almost convinced all these ppl getting huge views on here have either: -no personal lives -tech teams helping them run their setup -just plain lying Or I’m just retarded.

@juntao Bravo 👏👏

@RhysSullivan this flag should be the default... 🤣 I can always recover if it fucks up..

running claude code with --dangerously-skip-permissions is sudo coding

@trikcode Naaaa, we just start writing the code again... No biggie...

Dude imagine ChatGPT, Claude, and Gemini are taken down simultaneously We become cavemen again

@asaio87 Isn't that normal..? It's new, it has be learnt and set up. My bet is they will reap the benefits at some point, while those that don't bother and work a nice comfortable 9 to 5 get obsoleted...

All people I have seen using the so-called OpenClaw AI agents are going to bed every night at 2.30 AM Werent these agents supposed to give you more free time ? All these guys work 19 hours a day.

I've stopped trying to figure out the business model before I build the thing.. just create what should exist, talk about it, and see what happens. money is a lagging indicator. the building is the signal.

@corbin_braun I agree (but maybe maybe quicker)...? But I think distribution is already the only thing that matters.

I give it 2 years. before software is dead and anyone can instant build 90% of products. at this point. all that will matter is distribution and networks. adjust accordingly for end game.

I think this will be the first year we see unemployment climb in most countries while the economy still grows.

I think we'll see at least 10 million people running home-based AI agents on local hardware by the end of the year (eg. OpenClaw on a Mac Mini or Studio). Most people don't realise how close this is. The hardware is (almost) already there.. the software just caught up.

Sign post (adjacent): I just bought my first Mac Studio to run a basic model (with OpenClaw)... I expect Apple will start releasing AI optimized SKU's for this kind of stuff. eg. High unified memory and memory bandwidth. And maybe trimmed down everything else.. And maybe some out of the box distilled open weight models to run on them..

I’m now convinced that the biggest winner of the AI race will be Apple. They will acquire Anthropic and put an AI model that can run on ~32GB of RAM in every device. It will be private, local, have perfect memory, access to all of your files and it will cost $0. The real moat was owning the hardware.