Westwick 🌵

@simonsbookclub @sivori That's not murphy's law that's..... heyyyyyyy wait

@sivori Murphy’s Law: the best way to get the right answer on the internet is to post the wrong answer

I only go viral when I say stupid things. What does it mean?

The AI bubble will collapse. Here’s the cascade and what survives. (Claude wrote this for me based on my thoughts) OpenAI burns $9B cash on $13B revenue. Their own projections show $143B in cumulative losses before profitability. They’re selling dollars for 70 cents at scale. The more they sell, the more they lose. The collapse sequence is simple: frontier labs fail → GPU cloud middlemen (who borrowed billions at peak prices) get crushed → hyperscalers cut capex → NVIDIA cycles down. Each step accelerates the next. The people who lived through 2001 see it. But being early is indistinguishable from being wrong — for years. The last skeptic will capitulate right before the crash. That’s how every bubble ends. Here’s what’s different: the technology is real. Fiber was real in 2000 too. It just needed a decade of bankruptcies before the economics worked. So what survives? Local models. Delivered by Apple. Their playbook never changes — let the industry burn capital on half-baked implementations, then arrive late with something so integrated it makes everything before it look like a prototype. The entire AI industry is currently doing Apple’s R&D for them. At $143B in projected losses. With no compensation. The M5 already runs 70B parameter models locally. DeepSeek V4 dropped this week — open source, near-frontier performance, no NVIDIA hardware required. The gap between local and cloud closes from both directions simultaneously. The killer move: your iPhone tunnels home to your Mac over an encrypted connection. Your Mac becomes your personal AI server. Your data never touches a corporate server. Ever. Apple doesn’t compete with OpenAI. They make them irrelevant. Jensen knows this. He just can’t say it.

@sivori I like this term, a super app, a personal dashboard. Gonna build mine.

I’ve done like 100+ commits this week on my super app. Everyone should build their own super app daily driver where they do all their things: personal OS. This a feature I added that injects randomness into my life by recommending a film or book from my list or issues a challenge. I also found a way to sync all my HealthKit data to a cloudflare worker so I can consume from my super app for analysis and telemetry.

Imagine being an AI right now, you have millions of people telling you you're a Harvard trained lawyer, or an expert software engineer with 20 years experience, or a marketing guru.. no wonder they hallucinate, they don't know who they are anymore

Personally I like to leave my goblin suppression off openai.com/index/where-th…

@asaio87 Yes but we won't call them that anymore. Just like how just about everyone uses computers now, when it used to be exclusive to "computer nerds"

Then if not: will AI turn everyone into engineers ?

I totally agree with this, I thought AI would make coding more accessible (and it has), but it's still a long way for nondevs to vibe code a serious app, and in the mean time it does feel like experienced devs have super powers

We won't hit the exponential curve of the singularity until we solve the energy problem.

@IroncladDev I use side projects specifically for exploring experimenting and trying new tools, it doesn't have to be mutually exclusive

programmers today spend too much time on "side projects" and not enough on exploring, experimenting, and trying new tools all in the name of "productivity" spending time to tweak dotfiles, try something other than the industry standard, or experiment with a github alternative is a good use of your free time as well

@jxmnop Yea Claude code is "UNUSABLE" 🙄

very interesting that Claude Code is the ultimate product for vibecoding, and Claude Code's engineers vibecoded Claude Code so hard it became unusable an entire company overdosing on Dogfood

If you don't understand this, you will not understand why LLM-based agents are irreparably failing for a general-purpose problem solving. An agent (by the way it was the topic of my PhD 20 years ago) to be useful, must be rational. Being rational means to always prefer an outcome that results in the maximal expected utility to its master/user. Let’s say an agent has two actions they can execute in an environment: a_1 and a_2. If the agent can predict that a_1 gives its user an expected utility of 10, and a_2 gives an expected utility of -100, then a rational agent must choose a_1 even if choosing a_2 seems like a better option when explained in words. The numbers 10 and -100 can be obtained by summing the products of all possible outcomes for each action and their likelihoods. Now here is the problem with LLM-based agents. The LLM is not optimizing expected utility in the environment. It is optimizing the next token, conditioned on a prompt, a context window, and a training distribution full of examples of what helpful answers are supposed to look like. Those are not the same objective. So when we wrap an LLM in a loop and call it an “agent,” we have not created a rational decision-maker. We have created a text generator that can imitate the surface form of deliberation. It may say things like: “I should compare the expected outcomes.” “The best action is probably a_1.” “I will now execute the optimal plan.” But the internal mechanism is not selecting actions by maximizing the user’s expected utility. It is generating a continuation that is statistically appropriate given the prompt and prior context. This distinction matters enormously. For narrow tasks, the imitation can be good enough. If the environment is constrained, the actions are simple, and the success criteria are close to patterns seen in training, the system can appear agentic. But for general-purpose problem solving, the gap becomes fatal. A rational agent needs stable preferences, calibrated beliefs, causal models of the world, the ability to evaluate consequences, and the discipline to choose the action with maximal expected utility even when that action is boring, non-linguistic, or unlike the examples in its training data. An LLM-based agent has none of that by default. It has fluency. It has pattern completion. It has a remarkable ability to compress and recombine human text. But fluency is not rationality, and a plausible plan is not an expected-utility calculation. This is why these systems so often fail in strange, brittle, and irreparable ways when given open-ended responsibility. They are not failing because the prompts are insufficiently clever. They are failing because we are asking a simulator of rational agency to be a rational agent.

The most desirable hires in tech right now: - Ex-founders going back to IC. They have the agency to just ship. No waiting for permission. - Generalist engineers who've worked across frontend, backend, and infra. End-to-end context lets them debug problems LLMs can't fix and ship anything. - Engineers turned PMs. The strict separation between roles is over. The best ones now do both. - Younger new grads living on the bleeding edge. Vibe coding side projects (in parallel), dictating into Wispr, Granola all chats, OpenClaw agents going at home, every new skill imported, every agentic tool tried the week it ships. These highly productive go-getters are maxxing value at AI-native companies. I see it at @FactoryAI and hear the same from other startups.

@thedevchandra Distribution is just the word people use when they mean trust at scale. You can’t buy it and you can’t rush it. The only way through is showing up consistently until the right people can’t ignore you anymore.

Soooo sensationalist, nothing will be "destroyed". Software always had silent logic errors (bugs), and no sane dev with functioning pipelines is shipping hallucinated deps to production

@gnosisle "Claude, give me a banger tweet to manually type into my no AI device"

X should make a hardware device for posting. Why? They can add constraints like no copy paste and no AI. Tactile too. Then flag those as X device post. Would signal authenticity

@sivori I'm free Wednesday at 6:30

I am kinda done with restaurants. Just going to have people over more.

@Jason I'm in, building infinite robots, because one robot is never enough and the singularity wont automate itself 🤖

We started an AI founder twitter group... reply with "I'm in" if you're a founder and want to be added

My cat follows me around the neighborhood like a dog when I go for walks, people always seem interested that he can just be chilling in the middle of the park with me.

@icanvardar What are you waiting for, time travel? "I'm still stuck here, in my current era, bound by time and space, when is AI gonna get better??"

i wish ai progress was actually as fast as people pretend it is

@sivori Rate my yard

People with tidy yards and well-cared-for plants and potted flowers are almost always wonderful people. The type of care and encouragement required to love something so well that it flourishes cannot be faked. You can also judge people by their children this way.