verdverm

29 ms·

901.
by verdverm 4mo ago
https://trakkr.ai/bias
902.
by verdverm 4mo ago
I suspect this is a temporary optimization / exploitation until more people have better understanding of how to leverage Ai, though perhaps not and it is the result of (intentional) underinvestment in education.
903.
by verdverm 4mo ago
> I can imagine doing that Assume you meant "can't" here. I think you can see evidence of how much people are paying attention to the Ai code in their products. Google Cloud has become awful in the last year as they have o
904.
by verdverm 4mo ago
Thanks for more details. I've been getting into street photography lately and wanted to make renders of the images using old-school image processing techniques. There are a ton of them out there that could mix-n-match for some really c
905.
by verdverm 4mo ago
> From our conversations though, a lot of companies also don't want to lose complete control over the user experience As a user, I don't care, these companies have not been servicing us, nor have our best interests in mind. The
906.
by verdverm 4mo ago
I'll give massive kudos to whom ever over at Atlassian is actually listening. I've seen multiple feature requests / complaints dealt with in the last 1-2 years. For this one, I've been harping to them about how no one wa
907.
by verdverm 4mo ago
Most of the bigger open weight models are pretty good. You can get them per-token from companies like Fireworks or OpenCode
908.
by verdverm 4mo ago
I forget what podcast / guest, but I recall this phrase "inversion of control" when it comes to SaaS and the "it's on our roadmap" for every feature request and bug report. This was in the context of what a rea
909.
by verdverm 4mo ago
as always it depends on your hardware, the tiny models for embedding / reranking typically have low latency qmd is focussed on local to the point of designing around single machine setups and this creates a gap where one runs agents+qm
910.
by verdverm 4mo ago
yup, one of the reasons I built gmd to replace qmd, also that I wanted to use Typesense and incorporate llm-wiki concepts. Updating memory does depend on the nature of the change and may impact multiple memories or indexed files, and why ll
911.
by verdverm 4mo ago
This is a great point
912.
by verdverm 4mo ago
being close is still impressive, especially for their first (released) model gives me hope that the training moat is even smaller than we thought
913.
by verdverm 4mo ago
(-_- )╯ covid shots ( ^_^)σ testi shots
914.
by verdverm 4mo ago
Its not as good as GLM 5.2 for agentic workflows while also being bigger. Competition is going to be ruthless because the super low cost to switching. There is also AllenAi in the US, but they have yet to produce a model at this scale. Than
915.
by verdverm 4mo ago
If it's ~30% bigger and not as good as GLM 5.2, why would I tinker with this model? Maybe for the multi modal?
916.
by verdverm 4mo ago
I'm curious about the greedy algorithm vs something like compressed sensing. Can you expand on the algo choice and how it works?
917.
by verdverm 4mo ago
and their oligarch financiers
918.
by verdverm 4mo ago
indeed, I do support showing down and being responsible, forcing the companies to pay for electrical infra build out, using minimal water cooling systems they have all this money, they can afford adhering to stricter requirements
919.
by verdverm 4mo ago
1. did you look at the results, many of those are not launches at all, like #1 2. have you considered sentiment of the comments, upvotes are not always a show of support, I have seen many where the primary sentiment is how bad the show is 3
920.
by verdverm 4mo ago
The administration seems to actively avoid learning from what's going on in Ukraine. Russia was on the receiving end of US munitions precisely because cell phones gave precise location of gatherings.
921.
by verdverm 4mo ago
Stars were abused for spam, that's why it has been made private for the project owner eyes only. You can make a GitHub app and access a user's stars with their permission. It's not gatekeeping, you have to do more to be secur
922.
by verdverm 4mo ago
new accounts are restricted because of AI slop/spam, you'll have to wait something like a week or two
923.
by verdverm 4mo ago
you cannot transfer KV cache from one model to another
924.
by verdverm 4mo ago
If you switch models, you have no cache. There is no way to change this as they depend on sending the tokens through the weights and saving the computation results. Different model, different tokens, different weights, different calculation
925.
by verdverm 4mo ago
HN is not a jobs board, please see the Guidelines and FAQ at the bottom of most pages.
926.
by verdverm 4mo ago
We are entering the era of personalized software. I have something similar but I'm not trying to promote or make money with it. I somewhat prefer it stay off the radar so I don't feel the need to cater to other people's featu
927.
by verdverm 4mo ago
Outrage media economy / NIMBY? There is a general malaise at prices and a target is needed. Data centers and the Trump-Bibi-Iran War are currently those targets seen as the primary driver of high prices, which have consistently been th
928.
by verdverm 4mo ago
even more: https://hn.algolia.com/?dateRange=pastMonth&page=0&prefix=fa...
929.
by verdverm 4mo ago
Its unclear if its in vllm, the eval harness, or the weights. I'm running it the same as other qwen3.6 derived models and it appears to work, but other comments speak of waiting for support in their tool. Could be a chat template or co
930.
by verdverm 4mo ago
same with opencode, next on my evals.list are the fine tunes from big model traces, and then my own if ftuning looks reasonable cost/time wise I have two main tasks I want to see if I can improve, coding and doc understanding/summ

More ›