kgeist

11 ms·

151.
by kgeist 7mo ago
Probably because there's a ton of code that deals with nested parentheses across languages in the training data, and models have learned how to work around tokenization limitations, when it comes to parentheses.
152.
by kgeist 7mo ago
https://arxiv.org/abs/2603.01097 >Overall, our findings position LoRA as the complementary axis of memory alongside RAG and ICL, offering distinct advantages.
153.
by kgeist 7mo ago
Heh, for a couple last days, I've been doing this exact kind of "neuroanatomy" on Qwen2.5/Qwen3 too. Fascinating stuff. To make it easier to fiddle with the network, I created a small inference engine that is stripped of
154.
by kgeist 7mo ago
I had it the other way around when writing a JIT compiler for my toy language: it worked on Windows, but crashed on Linux. The compiler was written in C++ and used C++ exceptions. On Windows (MinGW), I enabled the LJSJ backend (longjmp&#x2F
155.
by kgeist 7mo ago
We (~200 devs) migrated from Jira to Youtrack 10 years ago, and its functionality has been more than enough. Honestly, I don't remember anyone ever seriously complaining about it, aside from maybe a few nitpicks. A very solid product.
156.
by kgeist 7mo ago
>The pages belonged to The Almagest, in which second century polymath Claudius Ptolemy described his vision of an Earth-centered cosmos. Where's the article wrong?
157.
by kgeist 7mo ago
I had something similar happen with skills today. A popup appeared saying, "hey, did you know ChatGPT has skills?" Clicking on it opened a new chat window, and after some thinking it said, "I tried to launch the built-in skil
158.
by kgeist 7mo ago
>Today, we’re releasing <..> GPT‑5.3 Instant >Today, we’re releasing GPT‑5.4 in ChatGPT (as GPT‑5.4 Thinking), >Note that there is not a model named GPT‑5.3 Thinking They held out for eight months without a confusing numberin
159.
by kgeist 8mo ago
Yeah, I turned it off completely after it once started generating random garbage in long conversations with memory enabled. Like, it would give completely unrelated replies: I'd ask about Go, and it would tell me about the weather in P
160.
by kgeist 8mo ago
Research [0] from Anthropic about juniors learning to code with AI/without: >the AI group averaged 50% on the quiz, compared to 67% in the hand-coding group And why would they do better? There's less incentive to learn because
161.
by kgeist 8mo ago
>I'm in my mid-40's now. I taught myself C when I was 15. I have no desire to use LLMs to pump out code. For me, it's the other way around: I'm glad that AI can write code for me. A few months ago, I moved from an eng
162.
by kgeist 8mo ago
I wonder, can't Amazon create a new legal entity to split AWS into "AWS-for-DoD" and "AWS-for-everyone-else"? So one can work with Anthropic and the other can't. Not sure how it works in the US.
163.
by kgeist 8mo ago
>but where do you think LLMs pick up these patterns? >LLMs are able to capably mimic the common patterns that how to write books have suggested for the last 100 years as ways to make your writing more "impactful" and attenti
164.
by kgeist 8mo ago
ASML, which is based in the Netherlands, produces chip-making machines which TSMC and everyone else use to produce said chips. I think they got some expertise too :)
165.
by kgeist 8mo ago
It's also fun when the official transliteration rules suddenly change: a visa/passport issued in one year has a different name in Latin than a passport issued in another year. I was once two separate people :)
166.
by kgeist 8mo ago
Heh, I had this exact thing when getting certified at Microsoft (remotely). They required me to enter my name exactly as it appears on my government ID (not a single Latin character), but their registration site... simply blocked any charac
167.
by kgeist 8mo ago
For logins, we're already used to the fact that they're expected to be in Latin. Having them in the native alphabet is more trouble than it's worth (one system supports it, another breaks etc., easier to remember one, in Lati
168.
by kgeist 8mo ago
The upfront cost isn't that big: a Go map+RW lock is probably a few hundred bytes. Allocating them costs far below 1 ms.
169.
by kgeist 8mo ago
Orcaman is a very straightforward implementation (just sharded RW locks and backing maps), but it limits the number of shards to a fixed 32. I wonder what the benchmarks would look like if the shard count were increased to 64, 128, etc.
170.
by kgeist 8mo ago
Were those 16 mln sessions used only for alignment, chat format, reasoning, etc.? Or it's possible to train a base model too? If a single session is at least 32k tokens, then it's already 0.5 trillion tokens to train on, interesti
171.
by kgeist 8mo ago
It's something we debated in our team: if there's an API that returns data based on filters, what's the better behavior if no filters are provided - return everything or return nothing? The consensus was that returning every
172.
by kgeist 8mo ago
I run Qwen3-32b locally without any tools (just llama.cpp) and it can do basic arithmetic for smaller numbers ( like 134566) but I didn't check it for much larger numbers. I'm not at the PC right now but trying to do it via OpenRo
173.
by kgeist 8mo ago
Not a security researcher, but I once found an open Redis port without auth on a large portal. Redis was used to cache all views, so one could technically modify any post and add malicious links, etc. I found the portal admin's email,
174.
by kgeist 8mo ago
>surprising amount of work Dunno, in my Go+HTMX project, it was pretty trivial to add SSE streaming. When you open a new chat tab, we load existing data from the DB and then HTMX initiates SSE streaming with a single tag. When the server
175.
by kgeist 8mo ago
8b models are great at converting unstructured data to a structured format. Say, you want to transcribe all your customer calls and get a list of issues they discussed most often. Currently with the larger models it takes me hours. A chatbo
176.
by kgeist 8mo ago
Maybe this: https://fil-c.org/pizlix >Pizlix is LFS (Linux From Scratch) 12.2 with some added components, where userland is compiled with Fil-C. This means you get the most memory safe Linux-like OS currently available.
177.
by kgeist 8mo ago
>For example, Go uses a non-copying GC because Google wanted it to work with their existing C++ code more easily. Copying GCs are hard to get 100% correct when you're dealing with an outside runtime that doesn't expect things t
178.
by kgeist 8mo ago
>What's dispiriting is the (lack of) process and care: take someone's carefully crafted work, run it through a machine to wash off the fingerprints, and ship it as your own. "Don't attribute to malice what can be adeq
179.
by kgeist 8mo ago
LLMs often already "know" the answer starting from the first output token and then emulate "reasoning" so that it appeared as if it came to the conclusion through logic. There's a bunch of papers on this topic. At l
180.
by kgeist 8mo ago
Here where I live a similar thing exists for like 10 years already, and if the recepient has 2 bank accounts from different banks, then the banking app lets you choose which bank you want to send the money to. They also show the recipient&#

More ›