Jeff Witters

@atmoio You’re surprised the masses can’t grasp nuance? Most of them can’t even spell it.

Lots to talk about today, including my recent unfortunate AI psychosis episode, OpenAI's leaked financials, and Stratechery's piece on the recent vibe shift towards Anthropic sympathizers. Timestamps: 0:00 - Psychosis 2:20 - OpenAI financials 5:50 - Stratechery piece on Anthropic 13:30 - The new less-funny era of AI

"He's too old." "He's lost his pace." "Leclerc is the better driver." Never write off greatness 💪

@TheAhmadOsman opensource & accessible for the average guy too. Not just you need a rack of Sparks or h100 to play.

Got online to dozens of emails from builders and investors on my Opensource AI Must Win declaration Apparently someone posted it on HN yesterday and it was the 2nd highest voted of the day Over the next few weeks I will be in discussions with researchers, investors, and others to ensure we bring that vision to life More soon

@BenjaminDEKR True but nuclear is missing the odd religious component that frontier AI elicits.

I am not saying this is the ideal model, but it seems very likely.

Nuclear power may be a good comparison to frontier AI here. There are non-government companies running nuke plants (GE, Westinghouse, Constellation, Duke etc) yet the government and NRC are always watching, and have the authority to stop anything they do not approve.

@gregisenberg Don’t you live in the US?

The takeaway from Fable 5 being BANNED by the government: GET GOOD AT LOCAL MODELS SO YOU HAVE 100% CONTROL. My entire weekend was going to be building my craziest ideas with Fable 5. That's now cancelled. So instead of building with Fable this weekend, I've decided I'll go deep on local models: 1. Start with the runtime. Download Ollama or LM Studio first. This is the thing that actually runs models on your machine. 2. Match the model to your hardware. A model's size is measured in billions of parameters (7B, 32B, 70B). Bigger is smarter but needs more memory. Rule of thumb: a 7B model runs on almost any laptop, a 32B needs a good Mac with 32GB+ RAM, a 70B needs serious hardware like a DGX Spark or a maxed-out Mac Studio. 3. Know which model for which job. Qwen 3 is the best all-around choice for most tasks. DeepSeek for reasoning and coding. Gemma 4 when you need something tiny that runs on a phone. Llama when you want the biggest community and the most fine-tunes. 4. Quantization. You can shrink a model to run on weaker hardware with barely any quality loss. Look for versions labeled Q4 or Q5. This is how a model that "needs" a server runs on your laptop. Learning this one concept changes everything. 5. Connect it to your agent. Point Hermes or your agent stack at a local model. 6. Context window is your real constraint locally. Cloud models give you huge context for free. Local models make you pay for it in memory. A bigger context window eats RAM fast. Keep your sessions tight and your prompts lean or your machine chokes. 7. Learn to give local models tools. A smaller local model with web search, file access, and code execution beats a giant model with none. The capability gap closes fast when you wire up the right tools. The model is the engine but the tools are the wheels. 8. Fine-tuning is more accessible than you think. You don't need this on day one, but know it exists. You can take an open model and train it on your own data so it gets good at your specific domain. I'll probably do a breakdown at some point on this @startupideaspod if people are into it. The lesson from this ban is basically don't build your entire workflow on something that can disappear with a single letter. Own part of your stack. Local models are insurance. It reminds me when people realized they don't own social media accounts. And then you saw people build email lists etc. I remember running a startup and my biggest traffic source was organic FB. All of a sudden, algo changed, and I lost 99% of my traffic. Same sorta moment (but bigger) for AI. This is a wake up call.

Pretty satisfied with my homelab / minilab at this point. 100% 3D printed Lab Rax

@atmoio Everything is possible now and I've never cared less about any of it. Both things true at once.

yesterday i signed up again for claude max $200 plan and had it change the whole visual metaphor of the productivity app i’ve been working on intermittently over the past year: instead of a traditional UI with tables, lists, tools, etc, i told Fable to use a desktop OS metaphor instead for displaying the various built-in mini apps (tasks, chat, notes, etc). all with a functioning dock and animated wallpaper and multiple window support etc. fable was able to solve the problem but really i’m beyond the point of being impressed by an LLM doing some upfront task. everything worked, it “made no mistakes”, all tests passed (it even fixed old tests), but i was like ok whatever thanks. i blew past my $200 limit in 2 hours. and now i’m sitting here like, ok, now what? do i ship this? hear me be a whiny bitch for a second: that it was too easy killed the whole part of the journey of making an app where you become a new person through the creation process, and you earn such pride in your work which in the past gave you the energy and courage to ship things. and i’m like, i can ship this. i can try to make a buck. the app is done. but i just don’t feel a bond with the work. now if you were a somewhat savvy operator, the business type that would happily sell refrigerator coolant if you sensed an opportunity, AI will be a godsend for you. but i don’t wanna sell refrigerator coolant. and now because everything is so easy, i hardly ever feel like i’m solving a real problem anymore. it’s like how deep of a problem am i really solving if someone can one shot my app in 2 hours? i will say that in those 2 hours yesterday, i really enjoyed being back near the code. there’s nothing funner than making shit. it’s just that the new way of doing things kills a lot of the creative and spiritual juices you used to get before, that many times lead to commercially beneficial outcomes. now, i just don’t know what’s worth building anymore.

@vanschneider “Everything is possible now and I've never cared less about any of it. Both things true at once.” That perfectly encapsulates design in 2026 so far. It’s like we finally got keys to the candy store and now hate sugar.

Spoke with many friends recently, designers and engineers, all people who've been doing it for 20 years or more and had a lot of success doing it. And they all say the same thing. The AI stuff is genuinely useful right now. It's fast and things that used to take a week take an afternoon. Things you never even attempted because there was no time, now you can just do them. It's the biggest enabler ever. But in the same breath, every single one also says that it's the least fun they've ever had in their entire career. They also mention it makes no sense to do it the old way. They're all in. It's a strange paradox which I feel myself. Everything is possible now and I've never cared less about any of it. Both things true at once. Not sure if thats just the feeling of the current moment, or if I just talked to people who're tired of the computer (since all of them been doing it for a long time).

@TheAhmadOsman One pricing change by the frontier labs and half these ex-crypto newly minted AI grifters are on to the next thing. We'll just be left with the slop.

Grifters shipping vibe-coded slop are everywhere now and it is getting exhausting ngl The issue is not that they are vibe coding - Build however you want - Use whatever tools you want - Ship fast, experiment, have fun The issue is pretending the output is serious software when it belongs in the trashbin A lot of these people are not building products - They are producing screenshots - GitHub activity - Fake momentum And naturally, these are the same folks walking around proudly showing their GitHub commit count as if “many tiny commits to a broken app” is a proxy for taste, architecture, reliability, or competence Very happy for them though: The graph is green, the software is not

Delete & Report Spam does absolutly zero in Messages

I struggle to understand the rush to a goal AGI. Most people seem perfectly content with AI as a tool while others talk about it like their quasi-religious destiny / goal is to birth a new species, and I’m not convinced those are even remotely the same project anymore. Somewhere the path diverged ... I don't know, its like the tools + the philosophy + ego is just getting overwhelming for what is a text file. I need to go for a walk.

not feeling the AGI

@zeeg Successful design isn't always original, and that's okay.

@jasonfried Miller’s Law + Cognitive Load. Problem with these interfaces they are trying to force feed is not everyone knows the question, or even how to ask the questions they need to ask, so they leave.

I think you're going to see it's all going to converge back to screens and data and panels and buttons. People don't want to ask the same question over and over. They'll ask something, it'll be set up to show something, and that thing will be saved as something they can always look at. Stable pre-defined glances, not blank slates each time. Common questions will become buttons and panels again. Most people ask the same kinds of questions about what they work on most of the time. Having to start from scratch with the questions every time seems like a step backwards. Another way to put this: Questions are wonderful for a deeper dive, but not a daily drive. Not sure you're suggesting questions always, but the comparison screenshots looked that way.

@Scobleizer @michaelfanous1 Privacy is dead & data breaches are on the horizon.

This guy has five million AI agents tracking everything publicly available about three billion people! He spent half a million bucks doing that. Be scared? Or be educated? @michaelfanous1 runs nyne.ai and I have a lengthy conversation about what he's doing and why it could lead to a new kind of social network.

@ThePrimeagen and birthed so many punchable faced bros that all they do is yell about Openclaw or Hermes

So this was the tweet that ruined it...

@adriankuleszo Design call-out culture, so stupid. Posts like this often make the accuser look worse than the person they’re accusing. ugh

I appreciate people looking out for other people’s work, but I don’t want this to turn into a witch hunt. We’re all inspired by hundreds of references daily. Every designer is. Nothing is fully original and nobody owns a layout. That’s how taste develops. Take inspiration and make it your own version. Whether it’s similar or not, I don't really care. We don’t need more drama in an already dramatic design space :D

@atmoio Man you really kicked the hornets nest with this one. Keep fighting the good fight 👏 👏👏👏

I'm done. I'm f***ing done.

@atmoio "sniiiiiiiifffffffffffffffffff, sniiiiiiiifffffffffffffffffff, sniiiiiiiifffffffffffffffffff"

Marc Andreessen accidentally told the truth about AI

@MilkRoadAI SPAC King Grifter and his shit sando's give idiotic take.

Chamath just delivered the clearest diagnosis of what is happening to enterprise software and the OpenAI Deployment Company is the most damning piece of evidence he could have picked. "The low end of the market is basically finished. There is no safe space." 90% of public SaaS stocks are down 30-80% from their 52 week highs, the median software stock is now negative over the last 3-6 months. Goldman Sachs reported that software forward P/E multiples fell from 35x to 20x, the lowest absolute level since 2014 and the smallest premium to the S&P 500 since 2010. The low end died first and fastest, because AI replaced it most directly. The small business tools, the lightweight project managers, the single function SaaS products that charged $49 a month per seat, those are being replaced by AI agents that do the same work as a workflow, not a product. You do not buy an AI powered tool, you describe what you need and it builds it and the seat based model that created the SaaS industry simply does not apply to that transaction. But Chamath's more interesting argument is about the high end and the tell he points to is perfect. OpenAI just raised $4 billion from 19 investors including TPG, Brookfield, Bain, and McKinsey to launch a consulting company and guaranteed those investors a 17.5% annual return to do it. On $4 billion in committed capital, that is roughly $700 million per year in guaranteed payouts, owed by a company that is projected to lose $14 billion in 2026. The goal of this venture is to compete directly with Deloitte, PwC, Ernst & Young, Andersen, and Cognizant. Think about what that structure reveals. OpenAI lost half of its enterprise LLM API market share from 50% to 25% between late 2023 and mid-2025, with Anthropic now leading at 32%. Its response was not to build a better model but rather to raise $4 billion, offer guaranteed PE-tier returns and hire embedded engineers to physically sit inside client organizations and make AI actually work in production. The reason, as Chamath identified, is that the high end of the market is not easy. "It's not like boop boop boop, put in a prompt and beep bap boop, it all works," he said and the data confirms exactly that. 88% of organizations running AI agents reported a security incident in the past year, 42% of C-suite executives say AI adoption is creating internal organizational conflict. The average enterprise AI consulting implementation costs $228,000 in year one versus $77,000 for platform-based approaches and most still stall before reaching production. Anthropic immediately matched OpenAI with a competing $1.5 billion consulting venture backed by Blackstone, Goldman Sachs, and Hellman & Friedman bringing the combined spend by the two leading AI labs on human powered enterprise deployment to $5.5 billion in a single month Chamath's read is that the high end, the large enterprise platforms like Salesforce with proprietary data flywheels, Palantir with its FDE model already proven at scale, Oracle with vertical specific data moats will survive and consolidate. The mid-market point solutions, the single function tools, the lightweight enterprise apps without defensible data assets, those are on the conveyor belt. The AI industry is not just disrupting the companies that use software but rather disrupting the companies that sell it.

@leopardracer Very similar setup minus the GPU. I run a 5090 32gb and run Zed. I use opus to plan and qwen to execute