jmalicki

7 ms·

91.
by jmalicki 2mo ago
Everyone knows steak is better well-done with ketchup.
92.
by jmalicki 2mo ago
It's a matter of cost. Did you see the 100x cheaper? If you have a workload that is going to be very heavy, incurring a large training cost to make a cheaper model work well with the dataset will be dramatic cost reduction. Most larg
93.
by jmalicki 2mo ago
I'm not saying it happened, I am saying they are working hard towards that goal as a business priority, and spending a lot of money on it. Software engineers are first, but other fields like finance and radiology have huge targets on t
94.
by jmalicki 2mo ago
I've actually been involved in annotation projects doing RLHF to train LLMs to do exactly that. It's not a matter of time, it's already happening - it's just seemingly lower priority than "profitable" projects
95.
by jmalicki 2mo ago
An idea can be both novel and derivative at the same time. You create idea B. It absolutely builds upon idea A, and is in some sense derivative. But your idea B introduces a lot of new novel concepts on top of idea A that makes it novel. An
96.
by jmalicki 2mo ago
All research builds upon previous ideas. E.g. you can patent a novel idea that would require licensing another patent to use.
97.
by jmalicki 2mo ago
I don't understand why from a legal perspective, but it seems like over the last decade or more the tech industry stopped caring about patents. Startups used to talk about their parents and large tech companies would encourage you to
98.
by jmalicki 2mo ago
Also, TIL about the Antonov AN-2 ("Wait... that kind of looks like a Cessna Caravan, but is a biplane, and is a tail dragger... what sort of aircraft is that?")
99.
by jmalicki 2mo ago
In California, it's because people (the state department of fish and wildlife) put the fish there. You can even see their scheduled fish plantings online: https://nrm.dfg.ca.gov/FishPlants/PublicPlantSearch
100.
by jmalicki 2mo ago
Isolated alpine lakes do not have connected rivers and streams, yet a ton are stocked with fish by humans. Is it unbelievable that ancient hunter-gatherers would do the same thing?
101.
by jmalicki 3mo ago
Yeah, true European cloud providers for these kinds of things seem to be behind, and a lot of the ones offering data compliance at the level of AWS are small enough that it's a bit harder to trust they'll be around and will keep t
102.
by jmalicki 3mo ago
> Worst-case example: Bootstrapped startup working in military. That's the easiest case. AWS Bedrock models running in AWS Secret Cloud for Industry. (I really have no affiliation with them, I'm just like... this is a complete
103.
by jmalicki 3mo ago
That's why you can define your own scope to control that.
104.
by jmalicki 3mo ago
A) is very doable with e.g. Amazon Bedrock. They'll give you HIPAA compliance, they even have a data center for US government classified data, they can give you European data sovereignty. And with OpenAI and Anthropic models to boot,
105.
by jmalicki 3mo ago
So you're talking about doctors in Europe. Europe has plenty of doctors.
106.
by jmalicki 3mo ago
This is the US. You can just throw a UFC party or something for Trump then you're good.
107.
by jmalicki 3mo ago
And he said "I'm not an expert" just after citing evidence.
108.
by jmalicki 3mo ago
A lot of people talk about building software "ecosystems". But that's more management/marketing than SWE, but that's where you build similar scale. You have Linux, products built on Linux, software written for Lin
109.
by jmalicki 3mo ago
But if you just hashed all the changes and put them in Merkle tree to verify, wouldn't that make it so much better?
110.
by jmalicki 3mo ago
No - inability to do so is part of the design of a good cryptographic hash, quite explicitly. https://en.wikipedia.org/wiki/Avalanche_effect
111.
by jmalicki 3mo ago
I don't know the exact tech behind it, but for a phone passkey I get a QR code on my laptop screen to scan with my phone, I accept it, and it logs me in.
112.
by jmalicki 3mo ago
You mean, it figured out how to drive to a solution to customer needs!
113.
by jmalicki 3mo ago
This sounds like it's for when you don't have a defined workflow and it helps you iterate on defining it by being deployed and identifying gaps.
114.
by jmalicki 3mo ago
Outcomes of leadership are hard to measure. If you pay by success of the company, noone is going to want to do a 5 year corporate turnaround.
115.
by jmalicki 3mo ago
> But you now have to trust that each department head is making good decisions. There is a lot of evidence that corporate mergers destroy value. What if... they were disintermediated by markets instead of being one company? There are ton
116.
by jmalicki 3mo ago
Oh yes, the agent won't have access to the weights via tool calls. But nothing would inherently stop an RLVR trained model from distilling a version of itself and proving it could regenerate that at runtime, if somehow it got off on an
117.
by jmalicki 3mo ago
> This is science fiction, these models don't have access to their own weights The weights plus the architecture is the model. What do you even think "the model" or "the weights" are? The weights aren't so
118.
by jmalicki 3mo ago
You can just tell your main agents "I have these implementation docs in <DIRECTORY>, please spawn parallel subagents to implement all of them" and I generally find that works, and what's great is it's in parallel t
119.
by jmalicki 3mo ago
I like to actually have hierarchical plans. Start with the big idea. Pin down the product manager-level description, and maybe some details about which features are in or out, how the phased design/roadmap will work, maybe what the ba
120.
by jmalicki 3mo ago
It's actually less token usage overall, since the gigantic context that builds up in the main agent is a huge driver of costs.

More ›