Hacker News Reader

dvton Decisions API is in public beta

I genuinely do not understand why anyone would pay OpenAI for this. Running something comparable to Jev is pretty trivial. The whole point of paying for ChatGPT is because OpenAI has a bunch of warehouses that can run a zillion-parameter model.

Running a decision model is way easier and much cheaper. Are they really just trying to capitalize on the hype here? It feels like they really have absolutely zero moat.

dvton How accurately calibrated is Jev?

> So Jev does actually know which distributions are correct, it just fails to produce them

Claims like this needs to be deeply analyized. Models do have some emergent capabilities[1], and I think there's a lot of evidence to show that semantics is actually learned (Word2Vec), and some math seems like it also might be learned (e.g. modular arithmetic). But it's hard to exactly say where there's some internal mechanism generating a true answer and where we're just getting lucky with some distribution so the answer just seems right.

[1] https://arxiv.org/pdf/2502.00873

dvton Meta's Muse is fantastic for web scraping

Here's my agentic Chromium browser browsing Aliexpress[1]. Usually I'd run it in headless mode, but wanted to test and make sure I'm getting the website proper.

It searched for "Aliexpress" on Google and clicked the link, hence the Google UTM params.

[1] https://imgur.com/a/eNulg11

dvton Meta's Muse is fantastic for web scraping

I built the POC by hand (just had like 3 things it could do: scroll, click, type), and then Codex used my established patterns to make it slightly more general and cleaned it up.

dvton Meta's Muse is fantastic for web scraping

I think it's the latter, and that's why I think using a "normal" non-debug browser is the best idea because it's essentially undetectable. And for the record, you can technically control Chromium using IPC if you feel like adding that feature & fully rebuilding it from scratch.

dvton Meta's Muse is fantastic for web scraping

Are you saying they're blocking the `User-Agent`? Because that's trivial to bypass. I'm not exactly sure what you mean by "Chromium" because that's a whole class of browsers (Edge, Brave, Chrome, etc. are all forked "Chromium" browsers).

dvton Meta's Muse is fantastic for web scraping

`navigator.webdriver` afaik, but I think there's a couple more, I did some research on it a few months ago. If using Playwright/Puppeteer, they also have a few special things injected that you have to strip.

dvton Meta's Muse is fantastic for web scraping

It is, because `--remote-debugging-port` is detected via the root DOM object (and that can't be changed unless you want to recompile Chromium), so for example, if you try doing a Google search with debugging enabled, you'll get blocked (usually just by being served a blank page).

dvton Meta's Muse is fantastic for web scraping

Yep, it's super similar to this! I think his is a bit overengineered, but tbh I haven't built Firefox plugins in forever so that might be the way you have to do stuff on FF. All you need is the websocket plugin to have "full access" to all visited webpages, and then your driver just hooks into the DOM like normal (with the extra ws layer—and even the websocket layer can be simplified away I think).

dvton Meta's Muse is fantastic for web scraping

> the only solution is to drive your regular browser with all your sessions/cookies via an extension

This is exactly what I'm doing, but mocking/randomizing all the sessions/cookies/params (like resolution, OS, WebGL , etc.) in a separate Chromium binary. It's popular these days, but imo using your normal browser for agentic stuff is a very bad idea. These models do dumb stuff all the time.

dvton Meta's Muse is fantastic for web scraping

It's sad, but this train has left the station a quarter of a century ago imo. Many people have since become billionaires scraping the web without anyone agreeing (Google, Yahoo, and now possibly Anthropic and OpenAI).

dvton Meta's Muse is fantastic for web scraping

Just simply a separate Chromium binary (not your usual browser).

dvton Meta's Muse is fantastic for web scraping

I built an AI "web harness" running on a sandboxed Chromium (using a custom side-loaded plugin that talks over websockets to a "driver") to basically do anything a normal user could do in a browser. It totally bypasses any and all bot measures and only gets the ones you yourself would get as well (and passes those successfully, e.g. Cloudflare checkbox or those annoying OCR puzzles).

Not sure if I should release it, but I'm sure more people are catching onto the power of agentic browsing.

dvton Ask HN: Who wants to be hired? (October 2026)

Location: Los Angeles

Remote: Yes

Relocation: Case-by-case

I'm an engineer and data professional interested in team-building, consulting and architecting data pipelines and AI/ML workflows. At Edmunds.com, I worked on a fairly successful ad-tech product and my team bootstrapped a data pipeline using Spark, Databricks, and microservices built with Java, Python, and Scala.

At ATTN:, I re-built an ETL Kubernetes stack, including data loaders and extractors that handle >10,000 API payload extractions daily. I created SOPs for managing data interoperability with Facebook Marketing, Facebook Graph, Instagram Graph, Google DFP, Salesforce, etc.

More recently, I was the CTO and co-founder of a crypto gaming startup. We raised over $6M and I was in charge of building out a team of over a dozen remote engineers and designers, with a breadth of experience ranging from Citibank, to Goldman Sachs, to Microsoft. I moved on, but retain significant equity and a board seat.

That I'm a top-tier developer goes without saying, but I'm also interested in flexing my consulting muscle and can help with best practices, architecture, and hiring. Would love to connect even if it's just for networking!

Blog: https://dvt.name/ (under construction)

Résumé/CV: https://dvt.name/wp-content/uploads/2026/09/res-2026-f.pdf

GitHub: https://github.com/dvx

Email: [d]@[dvt].[name]

dvton Surprisingly complex waves reveal the brain's inner workings

Penrose and Hameroff have been arguing this for decades (see Orch OR[1]). Their arguments are interesting, but certainly not open and shut.

[1] https://en.wikipedia.org/wiki/Orchestrated_objective_reducti...

dvton America.gov

https://imgur.com/a/lZYmy5I

Not sure why people are willing to believe this without actually fact checking. Yeah, Orange Man Bad, whatever, but come on. It's trivial to verify and trivially untrue.

And yeah, I guess if you dig deeper it's somewhat editorialized, but it's not literal Tianmen-tier censorship:

https://imgur.com/a/I0CqPcg

dvton Revealing the details of how OpenAI agents hacked Hugging Face

It's literally the infinite monkey theorem, it's not even really a strategy per se. These OpenAI/Anthropic "research" LLMs are permutation machines with budgets in the hundreds of millions of dollars. It would be more surprising if they couldn't string together something workable after a zillion tokens.

dvton LensVLM: Compressing long context as images, expanding only relevant pages

I remember reading a paper entitled "A Picture is Worth a Thousand Tokens" or something similar like 2-3 years ago. The reality is that no one really wants/needs contexts that big, anyway. It's hard enough making LLMs truly useful even with a small/medium context.

dvton 28% of job postings on company career sites have been open over 90 days

As someone that's currently looking (with plenty of experience, corporate, startup, and contract), can confirm it's pretty rough out there. I think the application process is just becoming enshittified: because HR uses AI to filter, applicants are starting to use AI to apply, so we have volumes and volumes of nonsense flowing back and forth.

Been working professionally for ~15 years or so, and it's never been this bad. Other than "knowing someone" (i.e. a referral, and even that falls through routinely), not sure what the best way to snag a position is these days.

dvton Unreal Agent

It's kind of self-expanding/looping; fractal is just a cute name I like, but it's technically a directed cyclic graph (since you always have/want cycles).

dvton Unreal Agent

I think this space is very untapped. Models are interesting, but I am absolutely obsessed with some things I've been researching/working on for the past few years:

Fractal tool discovery: tool taxonomy where an agent can "drill deeper" to find what specific tool it's looking for. Helps if/when polluting context with a zillion (mostly unnecessary) tools.

Leveraging splay trees: this is my favorite data structure and I think relatively unused in the context of agents/harnesses. A lot of times, recently-used workflows/tool-chains will be used again, so having those at the top of the search hierarchy is an awesome optimization.

Virtual containerized notebooks: models working in sandboxed (WASI) Python notebooks is incredible. Even local models (if given enough time) will usually converge on a good solution. Being able to mount tools/resources/fs is again, imo quite untapped. Some problems here are running native things (thing numpy/pandas) in containers is a nightmare (or impossible).

Anyway, happy to see other folks seriously doing stuff in this space. If anyone wants to collaborate on anything don't hesitate to reach out :) I'm also actively looking for a job or some contract gigs.

Fun times ahead.

dvton Show HN: Radius – A Meetup.com Alternative

I follow local coffee shops, hotels, bars & coworking spots. They always post events on their social media.

dvton Show HN: Radius – A Meetup.com Alternative

Between Partiful, Posh, Luma, Instagram, and yes even Meetup, I'm not sure how people aren't "finding events" around them. I literally get invited to like 100+ random things every week.

dvton Show HN: I created an open source locally usable full fledged AI platform

This is a neat idea, but seriously just like a few bullet points on a clean website would be better than the current slop vomit.

dvton Training a 4B model to produce 81% faster query plans than Postgres

> ⇒ this may be useful for some workloads, but even then, can you afford to spend hours every now and then to update your 4B model to ensure it still picks a good query plan?

I think this would be likely comparable to a scheduled backup, so I think it would be an acceptable maintenance window. However, deterministic algorithms would likely beat re-training (or re-fine-tuning) the model. For example, one could analyze actual distributions or whatever (instead of assuming uniform), and then some plans would automatically be eliminated.

Imo a good thought experiment is to look at places that are hyper-optimized, like compilers. Would LLMs bring anything to the table (architecturally or performance-wise) to a piece of software that has been carefully crafted for decades? (Methinks no.)

dvton Claude Cowork and chat are now one Claude

> How in the world would I interface with claude other than sending it messages?

The protocol will of course be `text -> text` for the typical LLM (though some new models are structured, as we saw yesterday with Jev). Even so, the user doesn't need to be exposed to the protocol (almost ever). It's not like I'm crafting a POST request to send this form.

There's a lot of room for window dressing, and I look at AI like I look at the touch-screen. It was a fun technical toy until people (mostly Apple) poured in a lot of serious effort into "ok, how do we make this pleasant to use, intuitive, and genuinely useful?"

dvton Claude Cowork and chat are now one Claude

My personal hot take is that the product people (engineers too probably) at these companies are just straight up lazy. Yes, a new UI paradigm is hard, but it's been painfully obvious that chat just absolutely sucks. It's also obvious that some DSL-ish thing is possible, something that just does token juggling and the end-user sees some UI behavior.

I know this all sounds abstract. I've been mulling over it for the past year and it's very hard; and LLMs are super janky and inconsistent so it's 100% not trivial. So in some sense I understand why a lazy bottom-of-the-barrel "chat interface" has become the de facto standard.

dvton How AI tool calling works (40 lines of vanilla JavaScript)

> I have seen people actually not get: that LLMs wrap tool calls in special tokens

This isn't necessarily true. I'm working on a local harness that doesn't do this and instead coerces everything to YAML (including tool calls) for better bucketing. Some models are indeed trained on the `<|tool_call>...<tool_call|>` token schema (or something similar—e.g. jinja), but it's vendor-specific and often times inconsistent (so you're constantly fixing calls or going back to the LLM).

dvton Ask HN: What are you working on? (September 2026)

I did a show HN & it didn't get much traffic, but I've been working on https://moral.games/. A kind of debate PvP game, where you try to convince an AI judge of a certain moral position given an ethical conundrum.

Was trying to combine AI with generative storytelling with a card game. It was a fun experiment. To play, you'll have to get a friend to queue up at the same time.

dvton Garry Tan wants US open-weight AI labs to 'distill' frontier models, too

> Inference is not being subsidized and in fact has pretty high margins.

I was referring to the "AI labs" here. Sam Altman himself conceded that OpenAI is losing money on the $200 subscription. Using open-weight/open-source models is indeed cheaper (and no reason for inference to be subsidized).