sroussey

6 ms·

91.
by sroussey 2mo ago
I think we are at the apex of Von Neumann machines. Once there is enough money to make alternatives, they will thrive. And AI is that catalyst.
92.
by sroussey 2mo ago
If you go for a full curve over the space, then yes. But like NAND, there is room for, I don’t know what to call it, “quantization”? You get a few values out instead of just binary. My guess is, like other issues of precision, errors can ge
93.
by sroussey 2mo ago
No, those doing this 100k times a day have very good data on this, good estimators and modeling. And the API has various knobs to change and evals will give you actionable data.
94.
by sroussey 2mo ago
And if at the border they found keys to your house so they went there and did a search too? Thanks what a phone is. Not even mentioning Tailscale. ;p
95.
by sroussey 2mo ago
Would that work in MySql? Back in the day it used to do things like "natural join" that just figured out the the common columns to join, but i can't remember if it did that with union.
96.
by sroussey 2mo ago
Ah, I remember those days when Clinton was in office and no more deficit. He also did a DOGE type thing to modernize and make the government more efficient (though in a measured way, not putting children in control). Those were the days.
97.
by sroussey 2mo ago
So many vacuous statements at the seam. This is the hermetic load bearing part, which I confirmed rather than assuming.
98.
by sroussey 2mo ago
I think around $300-400k, but that does not include the github issue/pr/ci stuff that runs continuously, which could double or triple it. Somewhere, the first two weeks were said to be about $200k.
99.
by sroussey 2mo ago
Oh, you could go analog rather than digital. Ever look at ALU design? Nothing efficient about it!
100.
by sroussey 2mo ago
It was a public demonstration to government procurement agencies.
101.
by sroussey 2mo ago
But power hungry. In that 5+ year timeline, the compute per watt could change by three orders of magnitude. GPUs are to LLMs what CPUs are to gaming — not a good fit.
102.
by sroussey 2mo ago
I can see this being real useful when they scrape the internet for content. And they don't want their own crap. Or just to see how prevalent their stuff is.
103.
by sroussey 2mo ago
It is worse than that, it can make text unintelligible. I have Claude explain what is happening in a PR, and the technobabble and use of rare words have to look up in a dictionary make it difficult to understand. I need another LLM to trans
104.
by sroussey 2mo ago
Oh, I have, and even Opus falls flat. I feel like regex would be better. Oh... maybe i can have Opus create such a regex...
105.
by sroussey 2mo ago
I have ubiquity UniFi for this reason (amongst others). OpenWRT is a good choice as well. Most home router software is such junk, might as well leave the door open..
106.
by sroussey 2mo ago
What’s good for data extraction? I have a hard time getting models to just pull names and titles from a blob of text.
107.
by sroussey 2mo ago
Most engineers have licenses in the engineering discipline, except Electric Engineers. Mechanical and Civil. Engineering is based on the constraints of physics thus applied to mankind uses. Now if you talk about software "engineers&quo
108.
by sroussey 2mo ago
The Euro AI Act basis for this is stupid. But from an AI scraper's point of view, it is great -- you can skip a lot of content that you yourself generated. This will have, as they say, unintended consequences.
109.
by sroussey 2mo ago
I tried tweeking for structured extraction, but got issues with token budget. What is the context size?
110.
by sroussey 2mo ago
Looking forward to npm version of needle-rs supporting v2. I added needle support for tool use in my side project.
111.
by sroussey 2mo ago
No, team is way too small. Team group is a better resolution. You want a certain amount of ownership for a repo that will survive a team being laid off, for example. And a bit of cross responsibility may seem annoying to a single IC, but is
112.
by sroussey 2mo ago
Our family computer when I was a kid was in my parents bedroom. There was only one one. Kids should have dumb phones not supercomputers in their pocket connected to everyone everywhere all at once.
113.
by sroussey 2mo ago
Each version minor version of Claude has its own preferences for vocabulary.
114.
by sroussey 2mo ago
One other advantage is item serial numbers. Or something else that makes an item that seems the same but actually be unique (perhaps the warehouse it’s in?)
115.
by sroussey 2mo ago
Or do a hybrid
116.
by sroussey 2mo ago
IMB YouTube video explainer: https://www.youtube.com/watch?v=ZfWDVO3rzeA
117.
by sroussey 2mo ago
Own the customer relationship.
118.
by sroussey 2mo ago
Instagram knows what content you pause on, which is a huge signal.
119.
by sroussey 2mo ago
Product search model: https://www.linkedin.com/posts/introducing-ontology-1-ugcPos... Edit: more direct links, sorry: https://onton.com/research/ontology-1 https://onton.com/resear
120.
by sroussey 2mo ago
15. Cut AI energy use by 1000x while increasing processing speed 1000x Obvious near term trillions dollar market to disrupt.

More ›