Alex Komoroske (@komorama) header
Alex Komoroske (@komorama) profile picture

Alex Komoroske

@komorama

Co-founder, CEO of @common_fabric Technologist obsessed with complex adaptive systems. https://nitter.cf/t.co/KAOw2Wxvrq

 Berkeley, CA
komoroske.com
Joined January 2008
  • Tweets1.3K
  • Following1.1K
  • Followers5.9K
38Media
Photo by Alex Komoroske (@komorama)Photo by Alex Komoroske (@komorama)Photo by Alex Komoroske (@komorama)Photo by Alex Komoroske (@komorama)Photo by Alex Komoroske (@komorama)Photo by Alex Komoroske (@komorama)Photo by Alex Komoroske (@komorama)Photo by Alex Komoroske (@komorama)
1/ It’s time to share more about what we’ve been working on: @common_fabric is a social computing lab building a medium for software that revolves around people, not apps. commonfabric.com
Image from Alex Komoroske (@komorama)
21
32
18
176
27,887
Different levels of agentic software have different risks today. 1. Vibecoding software that uses LLMs on your data. Some risk of prompt injection. 2. Sharing that software with your close friends. Now they can theoretically prompt inject it. If you'd trust them to be alone in your house for a few hours, it's probably fine. 3. That same software goes viral, and strangers you don't know can use it. A 1% chance of a successful attack, at a large enough scale, guarantees an attack. Level 3 is out of reach in today's physics of trust.
3
1
2
294
Netscape originally thought they'd be the next Windows. But the web was a fundamentally different kind of thing that enabled all sorts of newness -- think: publishing a page for anyone in the world to read. AI will be the same.
1
1
13
1,791
1/ How rocker.ai is more than an assistant; why we built a nervous system for companies: I spent years running the products behind 90%+ of Stripe’s revenue. At one point I counted 250 projects I was responsible for. My job was knowing the dependencies the teams themselves weren't aware of, which plans had to change, and what important historical context people were missing.
Rocker - The AI Assistant App for Work

Rocker - The AI Assistant App for Work

Rocker is the AI assistant for work that knows the history, the people, and what matters. For you or your whole company, connecting people and agents in flow.

rocker.ai
3
10
5
23
1,439
Little, helpful utilities are still dangerous. For example, a wordcount utility or file converter. You just hope the person who created the software won't misuse your data. You have to implicitly trust every piece of software you use, even for mundane tasks.
2
2
14
943
Great piece from Sam Arbesman: "When software finds its natural size, whatever results from it is a more natural form, perhaps even a more resonant kind of computing." arbesman.substack.com/p/soft…
Screenshot of a passage from Sam Arbesman's essay "Software Finds Its Natural Size": "The business of software is likely changing. Does it make sense to build a venture-backed startup in the software world? On the one hand, the trends of software democratization and vibecoding fit with the whole SaaSpocalypse: maybe bespoke programs can just be spun up at will? But it also seems that software produced by professionals and for a large audience of users can still be relevant. Making this software still requires craft and taste. But I think this is not the right thing to concentrate on. Instead we should be focusing on how to think about the very many scales and markets that are now available for different kinds of software. Specifically, it no longer means than every piece of software needs to raise venture capital funds, scaling to the moon. Rather, software can now find its natural size."

ALT Screenshot of a passage from Sam Arbesman's essay "Software Finds Its Natural Size": "The business of software is likely changing. Does it make sense to build a venture-backed startup in the software world? On the one hand, the trends of software democratization and vibecoding fit with the whole SaaSpocalypse: maybe bespoke programs can just be spun up at will? But it also seems that software produced by professionals and for a large audience of users can still be relevant. Making this software still requires craft and taste. But I think this is not the right thing to concentrate on. Instead we should be focusing on how to think about the very many scales and markets that are now available for different kinds of software. Specifically, it no longer means than every piece of software needs to raise venture capital funds, scaling to the moon. Rather, software can now find its natural size."

2
6
514
AI makes infinite software possible: if you want something to be different, you simply ask. Would you like your grocery list to be sorted by aisle? Ask. For programmers, that’s a massive reduction in friction. For non-programmers, it’s a revelation.
3
3
20
1,195
I just published my weekly reflections: docs.google.com/document/d/1… Tokens as opex vs. capex. Free-candy van. Agents as a school of fish. Agents as islands. Attention caches. Fail-small. Judgment vs output in an age of agents. Fractal coherence. Agent-assisted Strangler Fig rearchitectures. Heel-nipping proactivity. Hill climbing vs hill finding. Implicit time horizons.
Image from Alex Komoroske (@komorama)
1
3
23
1,383
The architecture of networked software has been a dumpster fire for 30+ years. It wasn’t that big of a deal because it was submerged. Now with Fable class models, it’s on the surface. The fix requires a fundamental rethink of the foundations. From @benthompson's Stratechery piece this morning:
Screenshot of a passage from Stratechery, under the heading “The End of the Internet.” It lists the past week’s security incidents: compromised certificate-authority keys, new vulnerabilities in Node.js, GitLab, Java and Cisco products, and breaches at LiteLLM and Railway. It argues that software rests on a human-built foundation whose main test was whether it worked, not whether it was secure. Attackers used to be limited by human time and skill; AI can keep trying without limit, so more holes are surfacing. It expects things to get worse before they get better, and predicts software will improve once AI rewrites it and agents defend it.

ALT Screenshot of a passage from Stratechery, under the heading “The End of the Internet.” It lists the past week’s security incidents: compromised certificate-authority keys, new vulnerabilities in Node.js, GitLab, Java and Cisco products, and breaches at LiteLLM and Railway. It argues that software rests on a human-built foundation whose main test was whether it worked, not whether it was secure. Attackers used to be limited by human time and skill; AI can keep trying without limit, so more holes are surfacing. It expects things to get worse before they get better, and predicts software will improve once AI rewrites it and agents defend it.

2
1
1
16
1,274
stratechery.com/2026/the-end…
The End of the Internet

The End of the Internet

The old Internet, where anyone could host a server is over, but it’s just the final blow in an ending that has been coming for years. The future, though, is brighter than it appears.

stratechery.com
143
It’s established that AI can do a lot of useful stuff for you… but only if it has the right data. The limiting factor is not how good the AI is, it’s trusting it enough to give it all the data it needs.
4
2
1
25
1,811
en.wikipedia.org/wiki/Copy_E…
114
It doesn’t matter if you wouldn’t be fooled, if your agent would be. Your agent will very helpfully read every word (including hidden text) and has no reliable way to differentiate between information and instructions. futurism.com/artificial-inte…
Using an AI Agent Makes You a Sucker for Phishing Emails

Using an AI Agent Makes You a Sucker for Phishing Emails

As cybersecurity software company Malwarebytes warns in a blog post, AI assistants may fall prey to prompt injection attacks.

futurism.com
3
472
Electricity was dangerous before it was convenient. Before it could be widely adopted, we had to find a way for electricity to fail small. (Think fuses and circuit breakers.) AI needs a way to fail small. A new @common_fabric lab note: commonfabric.com/notes/fail-…
Fail small to reach further

Fail small to reach further

In the current trust model of networked software, “more secure” means putting yourself in a smaller and smaller box. With a new trust model, “more secure” could let you reach further.

commonfabric.com
2
1
10
826
The difference between “you may” and “you must”: a platform vs a cage.
2
10
927
I just published my weekly reflections: docs.google.com/document/d/1… Software pyramids. AI-made sandcastles. Dead-end software. Spending tokens on opex vs capex. The Reset Nexus. Mentorship trust fall. The AI Peter Principle. The non-viable solution to Goodhart's law. AI making your spikes sharper. The cost of linearizing writing. The Lethal Trifecta Gordian Knot. Chats as a junk drawer.
Image from Alex Komoroske (@komorama)
3
11
5
140
34,226
I’ll be chatting with @timoreilly about our work at @common_fabric at 8am Pacific tomorrow. Hope to see you there! oreilly.com/live/live-with-t…
Image from Alex Komoroske (@komorama)
1
12
1,069
Is an agent just an extension of you and working under your direct supervision? Or can it take speculative actions on behalf of you? The latter is required to make it 1000x more powerful. But that requires a trust model where autonomous action is safe.
4
2
20
1,382
What would an LLM-native notification system for your life look like? Every app only cares about its own notifications, so no app can rank its content against the others’. Only a system dedicated to you can.
6
18
1,341
Agents can create a massive amount of coherence in a short period of time. And they can create decoherence just as fast.
1
10
633
We use “auth” to mean both authentication and authorization. That works when most actors are humans, who are embodied and can be held accountable. Agents break this norm. The difference will start becoming extremely important.
2
1
1
13
1,036
Load more