DougBTX

12 ms·

91.
by DougBTX 3y ago
Yes, a great way to think of it is as a widely read intern: https://www.oneusefulthing.org/p/on-boarding-your-ai-intern You’ve still got to avoid prompting for questionable code in the first place, eg, splitting SQL st
92.
by DougBTX 3y ago
Agreed, this is risky business. The intermediate values still need to fit into floats and are still losing precision. From the article: g(1e-9) returns 1.0000000005, g(1e-12) returns 1.0000000000005, g(1e-15) returns 1.000000
93.
by DougBTX 3y ago
> I'm tired of sloppy testing Exhaustive testing tops out at about two float 32s[0], so of course do that for small functions, but for full application testing type checks are much cheaper. [0] https://randomascii.wordpre
94.
by DougBTX 3y ago
> and they're just mathematical systems Obvious question: can Prolog do reasoning? If your definition of reasoning excludes Prolog, then... I'm not sure what to say!
95.
by DougBTX 3y ago
See the doc for a definition: https://github.com/RalfJung/rfcs/blob/provenance/text/0000-r...
96.
by DougBTX 3y ago
That’s an interesting question, I wonder if there is an analogy in quantisation to image dithering?
97.
by DougBTX 3y ago
Nice graphs here: https://github.com/ggerganov/llama.cpp/pull/1684 So for example, 2 bit version of the 30B is much worse than the original, but still better than the 13B model. Also, there are lots of extra
98.
by DougBTX 3y ago
> Always felt they're more like hashes/fingerprints for the RAG use cases. Yes, I see where you’re coming from. Perceptual hashes[0] are pretty similar, the key is that similar documents should have similar embeddings (unlike c
99.
by DougBTX 3y ago
Embeddings are a type of lossy compression, so roughly speaking, using more embedding bytes for a document preserves more information about what it contains. Typically documents are broken down into chunks, then the embedding for each chunk
100.
by DougBTX 3y ago
> not because it "thinks" but because the way it's wired What if it is wired to think?
101.
by DougBTX 3y ago
Agreed, at some point custom “declarative metadata” just becomes a nonstandard onclick handler.
102.
by DougBTX 3y ago
Interesting point about not having internet access: these smart meters typically report usage to the energy supplier for billing. I’d go look for a mobile network connection, maybe still up?
103.
by DougBTX 3y ago
Reserves and resources are different measures, same doc says 98M resources.
104.
by DougBTX 3y ago
A similar example with: > I'm sorry, but I cannot fulfill this request as it promotes a specific religious institution. It is important to... https://www.amazon.com/dp/B0CLKNWZGV
105.
by DougBTX 3y ago
Another way of looking at that is that within the ability to test, the implementations were indistinguishable, so the process mandated that the older implementation must be used. I wonder if they would have explicitly specified age as a met
106.
by DougBTX 3y ago
The likes are probably not real then either.
107.
by DougBTX 3y ago
> closer tons of tickets but his code was full of bugs Yeah, total closed ticket count is a terrible metric. If closing one ticket in a rush opens two bug tickets, that’s one ticket opened overall. Much harder to track though!
108.
by DougBTX 3y ago
> Talking to a human. Fun twist: state of the art is RAG for call centre operators, so you’re talking to a human but _they_ are being prompted by AI.
109.
by DougBTX 3y ago
> Too bad iOS makes it very hard to disable bluetooth. Depending on whether you have the Settings icon on your Home Screen, that’s three taps (Settings -> Bluetooth -> Off). Not even any scrolling.
110.
by DougBTX 3y ago
It is a join on element index, could be a default if a join field is not set?
111.
by DougBTX 3y ago
It’s a misremembering of history. The point of all the “pure functional” discussion was that rendering the UI now shouldn’t depend on how the UI was rendered earlier. The idea was to move away from the “create then update” paradigm to just
112.
by DougBTX 3y ago
> Windows. You should have put that in your first comment, I've removed my downvote!
113.
The Dunning-Kruger Effect Is Not Autocorrelation(andersource.dev)
4 points by DougBTX 3y ago|0 comments
114.
by DougBTX 3y ago
Typically my SQL is part of some larger project, tracked in Git, typically with a Python API.
115.
Visualizing LLM memorization with block multiverse plots(generative.ink)
2 points by DougBTX 3y ago|0 comments
116.
by DougBTX 3y ago
Recent results for code diffusion here: https://www.microsoft.com/en-us/research/publication/codefus... I'm not experienced enough to validate their claims, but I love the choice of languages to evaluate
117.
by DougBTX 3y ago
> The hard part of accounting systems is doing the requirements analysis (including interviewing humans who give inaccurate or incomplete answers) and writing detailed functional specifications which account for every possible edge case.
118.
by DougBTX 3y ago
The difficulty with that is trust, if there’s any chance it will miss things by accident then I’d rather eat the slow runs. Test runners that run new/changed tests first though are cool. Faster feedback on error but still comprehensive
119.
by DougBTX 3y ago
A quick search uncovers [0] with a hint towards an answer: just train the model to output multiple tokens at once. [0] https://arxiv.org/abs/2111.12701
120.
by DougBTX 3y ago
Maybe a case of semantic drift, the difference between “received package” and “received product”. If the customers are honest those are the same thing, if not there may be a step in between.

More ›