@tonyespinoza.bsky.social
Reposted
The problem with a dossier is that it's about you, not for you. It influences what results you get, but in a way you can't directly inspect or correct.
simonwillison.net/2025/May/21/...
simonwillison.net/2025/May/21/...
I really don’t like ChatGPT’s new memory dossier
Last month ChatGPT got a major upgrade. As far as I can tell the closest to an official announcement was this tweet from @OpenAI: Starting today [April 10th 2025], memory …
simonwillison.net
May 21, 2025 at 4:53 PM
The problem with a dossier is that it's about you, not for you. It influences what results you get, but in a way you can't directly inspect or correct.
simonwillison.net/2025/May/21/...
simonwillison.net/2025/May/21/...
Reposted
Another fun and wide-ranging discussion with @aish.bsky.social about last week's reflections!
share.descript.com/view/tGXfz0Q...
share.descript.com/view/tGXfz0Q...
bits and bobs week of 5/19/2025 - Descript
Hi, Alex,
Hi.
episode 11. I'm I was reading the bits and bobs and so many of them this week were from our conversation last week. But there are some very specific threads that I would like to continu
share.descript.com
May 26, 2025 at 6:44 AM
Another fun and wide-ranging discussion with @aish.bsky.social about last week's reflections!
share.descript.com/view/tGXfz0Q...
share.descript.com/view/tGXfz0Q...
Reposted
I just published my weekly reflections: docs.google.com/document/d/1...
Creating magic with LLMs' patience. Personal systems of record. Infinite software. Personal context engines. Emergent collective intelligence. Personal Tech. Swarms of gremlins. Sub-network metastasis.
Creating magic with LLMs' patience. Personal systems of record. Infinite software. Personal context engines. Emergent collective intelligence. Personal Tech. Swarms of gremlins. Sub-network metastasis.
komoroske.com/bits-and-bobs
Author : Alex Komoroske alex@komoroske.com What is this? During the week I take notes on ideas that catch my attention during conversations. Once a week I take a few hours to take a step back and try...
docs.google.com
May 26, 2025 at 7:13 AM
I just published my weekly reflections: docs.google.com/document/d/1...
Creating magic with LLMs' patience. Personal systems of record. Infinite software. Personal context engines. Emergent collective intelligence. Personal Tech. Swarms of gremlins. Sub-network metastasis.
Creating magic with LLMs' patience. Personal systems of record. Infinite software. Personal context engines. Emergent collective intelligence. Personal Tech. Swarms of gremlins. Sub-network metastasis.
Reposted
A security model that relies on LLMs to make security judgments on potentially malicious input is fundamentally unsound.
That's why prompt injection is such a hard problem to solve.
That's why prompt injection is such a hard problem to solve.
May 29, 2025 at 5:55 PM
A security model that relies on LLMs to make security judgments on potentially malicious input is fundamentally unsound.
That's why prompt injection is such a hard problem to solve.
That's why prompt injection is such a hard problem to solve.
Reposted
I want software that is person-centered, not origin-centered.
May 30, 2025 at 6:57 AM
I want software that is person-centered, not origin-centered.
Reposted
I just published my weekly reflections: docs.google.com/document/d/1....
Prompt injection. Master Control Program. Slopdev. LLMs as amplification algorithms. The limitations of chat-only UI. Coactive software. The context wars. The same origin policy as a human, not natural, law.
Prompt injection. Master Control Program. Slopdev. LLMs as amplification algorithms. The limitations of chat-only UI. Coactive software. The context wars. The same origin policy as a human, not natural, law.
komoroske.com/bits-and-bobs
Author : Alex Komoroske alex@komoroske.com What is this? During the week I take notes on ideas that catch my attention during conversations. Once a week I take a few hours to take a step back and try...
docs.google.com
June 2, 2025 at 2:19 PM
I just published my weekly reflections: docs.google.com/document/d/1....
Prompt injection. Master Control Program. Slopdev. LLMs as amplification algorithms. The limitations of chat-only UI. Coactive software. The context wars. The same origin policy as a human, not natural, law.
Prompt injection. Master Control Program. Slopdev. LLMs as amplification algorithms. The limitations of chat-only UI. Coactive software. The context wars. The same origin policy as a human, not natural, law.
Reposted
MCP is the AI era's OLE.
We’ve seen this movie before: new integration tech, huge promise, completely bonkers security assumptions.
We already know how this movie ends.
We’ve seen this movie before: new integration tech, huge promise, completely bonkers security assumptions.
We already know how this movie ends.
June 4, 2025 at 7:24 PM
MCP is the AI era's OLE.
We’ve seen this movie before: new integration tech, huge promise, completely bonkers security assumptions.
We already know how this movie ends.
We’ve seen this movie before: new integration tech, huge promise, completely bonkers security assumptions.
We already know how this movie ends.
Reposted
Big Tech’s AI Endgame Is Coming Into Focus
One app to rule them all
www.theatlantic.com
June 5, 2025 at 2:47 PM
Reposted
LLMs make infinite software feasible.
But if each particle of software is distributed as an origin/app, then the friction of orchestration dominates the value of the software.
Infinite software in the same origin paradigm doesn't fix aggregation, and might even accelerate it.
But if each particle of software is distributed as an origin/app, then the friction of orchestration dominates the value of the software.
Infinite software in the same origin paradigm doesn't fix aggregation, and might even accelerate it.
June 6, 2025 at 7:09 PM
LLMs make infinite software feasible.
But if each particle of software is distributed as an origin/app, then the friction of orchestration dominates the value of the software.
Infinite software in the same origin paradigm doesn't fix aggregation, and might even accelerate it.
But if each particle of software is distributed as an origin/app, then the friction of orchestration dominates the value of the software.
Infinite software in the same origin paradigm doesn't fix aggregation, and might even accelerate it.
Reposted
The user has to do less orchestration if they stay within one origin, so they do. Over time the origins with more data tend to get bigger and sprout more use cases, a form of data gravity.
This leads to hyper aggregation. That tendency is intrinsic to the same origin paradigm.
This leads to hyper aggregation. That tendency is intrinsic to the same origin paradigm.
June 6, 2025 at 7:09 PM
The user has to do less orchestration if they stay within one origin, so they do. Over time the origins with more data tend to get bigger and sprout more use cases, a form of data gravity.
This leads to hyper aggregation. That tendency is intrinsic to the same origin paradigm.
This leads to hyper aggregation. That tendency is intrinsic to the same origin paradigm.
Reposted
Infinite software won't counter the hyper-aggregation problem if the software is distributed within the same origin paradigm.
June 6, 2025 at 7:09 PM
Infinite software won't counter the hyper-aggregation problem if the software is distributed within the same origin paradigm.
Reposted
Sensitive data + tool use + LLMs + app-centric security model = danger.
LLMs turn any text into potentially executable instructions, exploding the attack surface of traditional security models.
LLMs turn any text into potentially executable instructions, exploding the attack surface of traditional security models.
Just blogged my hunch that the Apple intelligence Siri delay is because of security concerns around prompt injection here simonwillison.net/2025/Mar/8/d...
Apple Is Delaying the ‘More Personalized Siri’ Apple Intelligence Features
Apple told John Gruber (and other Apple press) this about the new "personalized" Siri: > It’s going to take us longer than we thought to deliver on these features and …
simonwillison.net
March 10, 2025 at 6:03 PM
Sensitive data + tool use + LLMs + app-centric security model = danger.
LLMs turn any text into potentially executable instructions, exploding the attack surface of traditional security models.
LLMs turn any text into potentially executable instructions, exploding the attack surface of traditional security models.
Reposted
Full comic here: smbc-comics.com/comic/strong-3
February 9, 2025 at 7:38 PM
Full comic here: smbc-comics.com/comic/strong-3
Reposted
I suspect AI will get smarter than most companies faster than AI will get smarter than the smartest people.
February 9, 2025 at 8:29 PM
I suspect AI will get smarter than most companies faster than AI will get smarter than the smartest people.