wongarsu

35 ms·

871.
by wongarsu 9mo ago
Which comes back around to logistics and scale: refining ore is cheap because ore is delivered on multiple 300t haul trucks or in giant trains
872.
by wongarsu 9mo ago
And it's worth noting that you can get DeepSeek at those prices from DeepSeek (Chinese), DeepInfra (US with Bulgarian founder), NovitaAI (US), AtlasCloud (US with Chinese founder), ParaSail (US), etc. There is no shortage of companies
873.
by wongarsu 9mo ago
Self-hosting training (or gaming) makes a lot of sense, and once you have the hardware self-hosting inference on it is an easy step. But if you have to factor in hardware costs self-hosting doesn't seem attractive. All the models I can
874.
by wongarsu 9mo ago
My rent is less than that. But if you add up salary, payroll taxes, benefits, social security etc my employer still spends around four times that amount on employing me. More if you include misc overheads associated with having one more emp
875.
by wongarsu 9mo ago
At my $work this manifests as more backlog items being ticked off, more one-off internal tooling, features (and tools) getting more bells-and-whistles and much more elaborate UI. Also some long-standing bugs being fixed by claude code. Head
876.
by wongarsu 9mo ago
The arguments for and against end up similar to those for and against banning drawn or AI generated depiction of csam. No actual children are harmed, it's artistic expression, moving the topic out of sight won't solve it, and any
877.
by wongarsu 9mo ago
Makes you wonder why there aren't more insurance companies out there using their data to lobby local governments to fix their road design. They have all the data to find hotspots, and reducing accidents would increase their margin. May
878.
by wongarsu 9mo ago
At first I thought it's a clever prompt because you see which direction the model takes it, and whether it "corrects" it to the more common "human riding a horse" similar to the full wine glass test. But if you tran
879.
by wongarsu 9mo ago
For reference, below is the prompt translated (with my highlighting of the part that matters). They did very much ask for this version of "horse riding a man", not the "horse sitting upright on a crawling human" version
880.
by wongarsu 9mo ago
They have image and video models that are nowhere near SOTA on prompt adherence or image editing but pretty good on the artistic side. They lean in on features like reference images so objects or characters have a consistent look, biasing t
881.
by wongarsu 9mo ago
Most of my PRs are drive-by PRs: I have an problem, maybe a bug or missing feature, that annoyed me enough to fix it. And because I want to use future versions without the work of maintaining a fork I instead invest the work to upstream the
882.
by wongarsu 9mo ago
Doesn't that just make it worse? It can't see color, but instead it can potentially see through thin clothing
883.
by wongarsu 9mo ago
> People only speak or write down information that needs to be added to a base "world model" that a listener or receiver already has Which companies try to address with image, video and 3d world capabilities, to add that missin
884.
by wongarsu 9mo ago
The same insights still hold true for streets and paths. Of course a single human or even bicycle can move with fewer constraints than a car, but a stream of humans won't. When we design pedestrian infrastructure with sharp corners peo
885.
by wongarsu 9mo ago
If we focus on base models and ignore the tuning steps after that, then LLMs are "just" a token predictor. But we know that pure statistical models aren't very good at this. After all we tried for decades to get Markov chains
886.
by wongarsu 9mo ago
It's incredibly ironic how accessibility is about going the extra mile to make things more accessible, and then people turn it into a weird acronym that only insiders understand, in an effort to avoid typing another 9 letters
887.
by wongarsu 9mo ago
It seems to have stabilized around $70k. Obviously after it had "stabilized" at $90k for two months, so it might drop further. But if I had to submit a bet I would guess it will fluctuate around the current level for another coupl
888.
by wongarsu 9mo ago
From what I gather the conversion wasn't a big deal. The engines of the time weren't picky about fuel, so you just have to find space to mount the wood gas generator (a very simple if bulky device) and pipe the wood gas into the f
889.
by wongarsu 9mo ago
The windows kernel is great. It's the stuff built on top of that that sucks. I doubt MS puts the best of the best on coding the start menu
890.
by wongarsu 9mo ago
If it was only network effect, then how did TikTok grow in a space where Instagram and Youtube were already much bigger players? How did they gain that user base? Network effect helps, but it only explains why they stay big, not how they go
891.
by wongarsu 9mo ago
But go just one layer deeper to 'why is every teenager using Tiktok' and the primary answer once again becomes 'Tiktok's recommendation engine'
892.
by wongarsu 9mo ago
I'm willing to bet "chip optimization work" doesn't mean "the work required to optimize a chip", but "some work tasks performed as part of chip optimization". Basically they sped up some unknown subse
893.
by wongarsu 9mo ago
Also choosing which methodology is the "right" one to measure a specific number. There are lots of ways to measure ethnic groups, the size of the capital or the unemployment rate. If you publish the numbers you get to choose which
894.
by wongarsu 9mo ago
Most countries have some kind of statistics department that publishes that kind of data in great detail. The issues start when you try to compare data, because different sources will use different methodologies
895.
by wongarsu 9mo ago
It's unloved on Linux because using Linux under memory pressure sucks. But that's not a good reason to abandon improvements. Even more so with the direction RAM prices are headed
896.
by wongarsu 9mo ago
I spend most work days with robot dev software (claude code). If Copilot had a similar ability to do useful work with meaningful oversight I wouldn't mind spending time with it. Sadly it does not
897.
by wongarsu 9mo ago
I've also made decent experiences with continue, at least for autocomplete. The UI wants you to set up an account, but you can just ignore that and configure ollama in the config file For a full claude code replacement I'd go with
898.
by wongarsu 9mo ago
I have a hall-effect keyboard from Wooting, and they are indeed excellent. Very reliable, and setting the trigger point in software/firmware allows a number of interesting features like triggering different key codes depending on how f
899.
by wongarsu 9mo ago
> But this administration has a limited shelf life, even hypothetically just under 3 years of immunity isn’t enough for comfort. Depends on how much faith you have in the current administration. Russia limits presidents to two 6-year ter
900.
by wongarsu 9mo ago
Now I wonder how strong the correlation between coding performance and ao3 knowledge is in human programmers. Maybe we are on to something here /s

More ›