Trying to figure out how search works
- Tweets 21,377
- Following 837
- Followers 4,287
- Likes 20,607
Slides protip: instead of a slide full of bullets, make each bullet its own slide and draw a picture to illustrate this point
1
2
7
666
Enthusiasm in a domain will have a far higher impact to your career than trying to "be an expert".
To demonstrate enthusiasm don't obsess over novel research, communicate thoughtfully what you're learning, from a newb perspective. And never lose the newb perspective
4
1
31
1,292
Expertise is its own trap, closing you down to new ideas.
Cultivate Shoshin, start each day assuming you know nothing about your domain
en.wikipedia.org/wiki/Shoshi…
2
3
226
I think what slop-debates show is that in most domains where AI is being applied, you can be 10x as valuable as the person putting in no effort by just putting in very minimal effort
1
6
452
Indexing your own traces is a great way to get introduced to @turbopuffer
piped.video/watch?v=ht0sxHBm…
1
22
1,644
I tweeted this earlier this week, and @turbopuffer and @perplexity_ai seem to have responded with contextual embedding goodness :)
The emphasis on chunking as a *retrieval* strategy for single-vector search has long been misguided.
You can't a-priori anticipate every query some text might answer. Then try to focus in on that bit of text (cutting off information). And a single vector obviously can't capture pages and pages of information
3
1
15
3,246
It's true agents probably write better code than the average human
The flipside is they also gaslight much better too :-/
4
460
Everyone should index their coding agent traces. TPuf is a pretty ideal sink.
maven.com/p/e48f16/high-scal…
2
2
30
2,597
In RAG, people spend too much time chunking, and not enough time working within the logical structure of their data
Agents understand book -> chapter -> section.
They don't get random 256 token seems
4
43
3,581
RIP Vector Database from @turbopuffer
turbopuffer.com/blog/rip-vec…
7
40
594
77,421
Personally, I have been a principal engineer for 5-10 years. And my job honestly hasn't changed (as much) with AI coding.
I delegated most coding to the team, got big ideas in place, and occasionally tinkered / sketched in code to get something on the right track.
1
13
1,469
I wouldn't bank your career on Jevon's paradox
Demand for SWE as a skill can increase. But the value of that skill can still go down.
IE a lot of people write. Being a good writer is useful. Very few people make tons of money just being a "writer".
4
4
96
8,425
Hey, I've got a talk using @turbopuffer to build agentic memory tomorrow. Come hang out? Learn a bit about memory? See what I've been up to :)
maven.com/p/e48f16/high-scal…
2
13
777
Doug Turnbull retweeted
I think memory is just a very poor framing and it's all just context. Sometimes that's context about user preference, or facts stored from earlier in the session and other times it's from static/dynamic files creating or uploaded to the session. It's all just context and information. Giving it more labels than that just gets messy and doesn't really change anything. Unless it's in the weights it's all retrieval.
2
1
3
162