Hacker News: wolttam

New comment by wolttam in "GLM-5.3-Flash"

wolttam — Wed, 26 Aug 2026 18:15:56 +0000

That will hurt latency and latency is very important for good tensor-parallelism performance

New comment by wolttam in "Qwen3.8-Flash-Next"

wolttam — Wed, 26 Aug 2026 16:45:43 +0000

It depends how you use the models. These small models work great for developers who prefer to stay more in the loop, and only task the model with things that can really only be interpreted in one way.

Not to mention, they’re great for self-hosting and getting yourself to not be dependent on some API that can go down or be altered at any time.

Big models seem to mostly be good for pushing ahead the frontier - the smaller models tend to gain the frontier’s capabilities after only a handful of months anyway. Many are perfectly content remaining a few months behind the bleeding edge.

New comment by wolttam in "GLM-5.3-Flash"

wolttam — Wed, 26 Aug 2026 16:25:14 +0000

Hopefully you also bought a switch

New comment by wolttam in "GLM-5.3-Flash"

wolttam — Wed, 26 Aug 2026 14:38:00 +0000

The recent and slightly smaller DSv4 Flash is also GLM 5.2 equivalent (or close enough)

New comment by wolttam in "New MCP Roadmap"

wolttam — Sat, 22 Aug 2026 15:59:06 +0000

The model has zero awareness of MCP, it’s the harness’ job to talk to the MCP server and simply present the model with the tools just like any other tool. The only giveaway to the model about where the tools come from is the ‘mcp__’ prefix in the name

New comment by wolttam in "I'm Becoming AI-Blind"

wolttam — Fri, 21 Aug 2026 14:35:32 +0000

I think I might have been a bit heavy handed in my comment as I was rebutting the idea that only tech people can tell. I suspect it helps to have been exposed to a lot of earlier model writing, which was even more sloppy and had more of the kinds of tells we still see today.

And I’ll concede on both ends that there are probably times I suspect content is AI generated when it isn’t, and times I suspect it isn’t generated, but it was.

AI tells seem inevitable. You have millions of people communicating with one effective “personality” that has tendencies to write in certain ways. If its content is published verbatim, then it will be easier to tell whether some content is AI generated just based on its similarity (sharing certain linguistic features) to other content being posted.

It’ll never be black and white though.

New comment by wolttam in "DeepSeek-v4-flash-vision-exp"

wolttam — Fri, 21 Aug 2026 14:22:12 +0000

Neither. Performance of all models is incredibly spikey.

New comment by wolttam in "I'm becoming AI-blind"

wolttam — Fri, 21 Aug 2026 14:05:38 +0000

Certainly not. Anybody who cares about language to any reasonable degree surely notices and is repulsed by heavily AI generated content.

New comment by wolttam in "DeepSeek-v4-flash-vision-exp"

wolttam — Fri, 21 Aug 2026 13:53:33 +0000

No one’s mentioned robots, so… robots. VLA models, etc.

New comment by wolttam in "US debt tops $40T as Treasury doubles bond buybacks to calm markets"

wolttam — Thu, 20 Aug 2026 19:59:44 +0000

Go ask a (capable, preferably open) LLM. If it doesn’t give you a direct answer it will give you enough of a starting point to start asking more questions.

(This is not financial advice)

New comment by wolttam in "Scientific study reveals TikTok videos deactivate key cognitive brain regions"

wolttam — Thu, 20 Aug 2026 19:47:58 +0000

They are different categories. Brain rot isn’t an imagined thing and we aren’t confused about where it comes from

New comment by wolttam in "Slack Code"

wolttam — Thu, 20 Aug 2026 15:25:26 +0000

This is probably getting close to the right interface for these things, and that's coming from someone who is religious about using terminals, neovim, etc.

New comment by wolttam in "fx :Tiny, open, native coding agent."

wolttam — Thu, 20 Aug 2026 06:40:17 +0000

It’s also exactly what the tech enables. It’s not hard to imagine there being hundreds of thousands of different harness projects, if not more.

New comment by wolttam in "Asana cleared 5 years of engineering work in 2 weeks with Codex"

wolttam — Thu, 20 Aug 2026 06:30:37 +0000

Pretty big self-own

New comment by wolttam in "PostgreSQL for Everything"

wolttam — Wed, 19 Aug 2026 14:45:20 +0000

That's just it - most use-cases are pretty basic, and if you don’t know what you need then Postgres is probably a great place to start.

If you’re just starting out, keep things simple. Otherwise, you probably already know exactly why you need something more than Postgres.

New comment by wolttam in "The Amazon Tax"

wolttam — Tue, 18 Aug 2026 14:34:35 +0000

I would have thought the Amazon tax is “free shipping” that has altered the pricing of every item, and now B&M retail is picking up the same pricing despite a much lower shipping overhead

Ads too, I guess

New comment by wolttam in "Incident with Github.com"

wolttam — Mon, 17 Aug 2026 19:53:24 +0000

+ issues + pull requests. There’s some nice to haves in there, that you’d need some other service or convention to manage

New comment by wolttam in "I checked 30 frontier model cards. Here are the benchmarks labs report"

wolttam — Sun, 16 Aug 2026 04:28:52 +0000

I’m getting good at reading LLM-speak (which is probably sad)

It means two “layers” of information access/validation.

Sounds like it first made a little map of model <-> benchmark, then went and filled in the score boxes.

Definitely not LLM layers

New comment by wolttam in "I Remain a Skeptic"

wolttam — Sun, 16 Aug 2026 00:14:09 +0000

I think an argument can be made that some software has gotten more secure.

I don’t think it can be denied that the models do show an ability to find security vulnerabilities that may have otherwise been missed

New comment by wolttam in "GenRec: Towards LLM-Native Recommendation at Netflix"

wolttam — Sat, 15 Aug 2026 13:41:48 +0000

It feels icky to me that they feed a log of what you’ve watched, when, and on what device, over to an LLM just to get back some recommendations.

Even though It’s probably exactly the type of thing I would do on a locally hosted LLM