designing interfaces, testing agents, building strange little tools
- Tweets 3,600
- Following 287
- Followers 418
- Likes 1,386
Pinned Tweet
Sunlight
We get it every day (even if diffused by clouds) but do we ever stop to think about how dependent on it we actually are?
Opus 5.5 helps visualize it. Sound on pls
4
6
15
143,630
Day 12 of diving deeper into AI Engineering
Memory & context (how an AI "remembers")
Here's the secret: the model remembers nothing. It only sees what's in its context window right now. Anything that looks like memory is a system you built around it
The context window (short-term):
- the window: everything the model can see, right now
- your prompt: plus the conversation so far, all fed in as text
- token limit: the window fills up, and old stuff falls out
- forgotten: scroll up far enough and it's just gone
Long-term memory (you build it):
- save it outside: notes stored in a database, not in the model
- summarize: squish long chats down to the gist
- retrieve: pull the right notes back when they're needed
- feed it back: drop them into the window, and the model "remembers"
The model itself remembers nothing. Memory is a system you build around it
1
3
221
🚨BREAKING: I asked Claude to upgrade my LinkedIn profile.
It didn’t just “upgrade” it. It turned it into a recruiter magnet.
Here are the exact 7 prompts I used:
2
2
294
If if there’s anything as disappointing at watching a fascinating video only to realize it’s ai
Like at least don’t make it realistic
2
4
347
Who ever thought Sam A would be the voice of reason in this??
As much as I don’t like him and some of the stuff OpenAI does, I really don’t want it to go under. It’s the only company capable of competing with Anthropic right now and we need the competition and fighting they bring
And also this 👇🏻
I am very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models, and think it is a real safety issue.
2
7
305
Corporate drama + dating drama + ai releases + 💩posting + y (insert yours here) = X (formerly Twitter)
2
5
136
Day 11 of diving deeper into AI Engineering
Guardrails (the checks keeping your AI in bounds)
A model on its own will try to answer almost anything. Guardrails are the checks you wrap around it, on the way in and on the way out, so it stays safe and on-topic
On the way in:
- prompt in: the user's request arrives
- input checks: scan for jailbreaks and prompt injections
- redact: strip personal data before the model ever sees it
- pass or block: clean requests go through, bad ones get stopped
On the way out:
- model answers: a draft reply is generated
- output checks: scan it for leaks, toxic content, made-up facts
- fix or replace: edit it, or swap in a safe reply
- safe answer out: the user gets a clean reply
Guardrails don't make the model smarter. They make it safe to ship
4
80
Finish the rhyme (wrong answers encouraged)
Beans, beans, the musical fruit,
The more you eat, the more you
2
2
108
