HNHacker News
TopNewBestAskShowJobs

nostrebored

2,711 karma · joined March 20, 2015

Opinions discussed here are my own.
submissionscomments
nostrebored··on Inkling: Our Open-Weights Model
it is pretty good at instruction following and has extremely fast decode.
nostrebored··on Bonsai 27B: A 27B-Class model that runs on a phone
qwen3.6 starts at 27B
nostrebored··on I used Claude Code to get a second opinion on my MRI
Oh I agree with you, it's just that I don't think LLMs are either. If you think of LLM knowledge, especially in scientific/engineering fields, as a lossy representation of the density of ideas, then you'd expect to see some weird behavior. I'm sure there is some sort of a temporal discounting and people thinking about this, but a naive NLL or Reverse-KL on medical literature would engrain some weird, wrong ideas.
nostrebored··on I used Claude Code to get a second opinion on my MRI
Daily allergy tablets are associated with huge increases in early onset Alzheimer’s. Glad you found something that works, but might be good to get some of the allergen injections :)
nostrebored··on I used Claude Code to get a second opinion on my MRI
I don’t understand the negative reactions. Medical care as it exists requires the doctor and patient to have their brains switched on. I’ve almost never had a problem where a doctor provides me with a diagnosis and I go about my day. Most of the times that I have, I’ve been confident about the problem and known what I needed. The doctor was a barrier to accessing care.

Dr. GPT is a good brainstorming tool. It helps synthesize information in a way that primary texts don’t. But it does force you to say “that doesn’t make sense”.

I do think that people saying “doctors don’t know the state of the art” have a weaker case. If you think about it in terms of token density during pretraining and how post training datasets are constructed, I think it would take us a very long time to adapt to any fundamental shifts. If we have forgotten how to cure scurvy, how many journal articles would it take before we adapt to a discovery?

nostrebored··on I used Claude Code to get a second opinion on my MRI
Claude is the worst FM at image understanding. Prior to gpt-5.4 the only usable models were Gemini and Qwen.
nostrebored··on I used Claude Code to get a second opinion on my MRI
But the binary classification models can be made ternary easily. RL on congruence plus penalty for misdiagnosis is easy to set up and gives great results.

What I’ve seen be the true bottleneck is people not setting up the structured data. But making a tiny reasoning model with OPSD -> GRPO is totally doable with a bit of money.

nostrebored··on I used Claude Code to get a second opinion on my MRI
I think that much of the visual gap is because what to attend to in images is less structured. Anecdotally small qwen finetunes (ie less than 10B) take task accuracy from sub 30% on FMs to 90%. We have sold some of these for outcome based back office tasks.

I think we’ll see a lot of specialized VLMs that provide real value.

nostrebored··on U.S. government will decide who gets to use GPT-5.6
He made it happen by continuously using doomsday marketing to pump up model capabilities. This is the comeuppance.

There is a huge contingent of people who do not interact with AI on a daily basis. Many of them legitimately believe we're seconds away from wiping away white collar America. Many more believe that we are creating the literal singularity.

None of them have seen claude try to model a problem they're familiar with.

nostrebored··on Qwen-AgentWorld: Language World Models for General Agents
Well if you can do this then you don't delegate execution path derivation to the agent. The benefit is a predictable coherent world state where you understand the impact of { current state } x { action } without having to enumerate that huge cartesian product.
nostrebored··on When I reject AI code even if it works
The leakage problem is so pervasive. None of the frontier models seem to have any idea how to actually hold out rows. God help you if you decide to change the data mix.

I was working on creating a next-n-actions predictor for one of our use cases and not paying much attention for a PoC. I was fairly happy with the progress for a few days, before actually reading the eval code and seeing that we leaked the final state in every eval.

It's nice to let claude run loose on porting from framework to framework (port my code from TRL to NemoRL to Tinker to VeRL) but looking at what it does in the intermediate steps makes me want to claw my eyes out. And getting it to adhere to our domain model (e.g. we have an SFTConfig and a .to_trl(), or a Row and a .to_harmony()) is impossible.

nostrebored··on Agentic Resource Discovery Specification
Building an ontology of how people think of and organize information, processes, and actions is not solved via markdown. It’s not well solved intra company much less inter. The systems that do solve some of this are optimizing for unstructured retrieval and will continue to.

Imposing a structure on this unstructured discovery feels like it ignores all progress in IR, and imposes a farcical structure that doesn’t have a tangible benefit.

nostrebored··on Claude: Elevated errors across many models [resolved]
I strongly prefer codex. Claude is annoying. Codex provides descriptions where I want them and more touchpoints to audit the quality of work. Claude code on experimental seems to not even show diffs when asked anymore, and it's much less clear what is being shipped.
nostrebored··on Claude: Elevated errors across many models
But how many plugins are people actually using? I can think of one MCP server I find valuable (context7) and one plugin that i've installed, but continuously think about uninstalling (obra/superpowers).

Both were trivial to set up with codex.

nostrebored··on /architect: Reduce Fable tokens by 80%, Fable orchestrates/reviews, Codex builds
LLMs are obsessed with “gates”. Freezing the gates here is intuitive to me as this point — don’t let validation drift.
nostrebored··on /architect: Reduce Fable tokens by 80%, Fable orchestrates/reviews, Codex builds
But why does your agent control doneness? It seems to me the most odd part to delegate. All LLMs are terrible at it. Most LLM tasks can be expressed as a DAG or DAG of DAGs. Why delegate that to a random point in context instead of enforcing the flow?
nostrebored··on New U.S. college grads now have higher unemployment than the average worker
i think the reason for your deluge of downvotes is that a society the promotes more things becoming affordable is one that prioritizes stability. universal access to housing, healthcare, and education that people want is only possible in a society that is immensely productive.
nostrebored··on When AI Builds Itself: Our progress toward recursive self-improvement
Do you not think that the allocation of human time is one of the world’s biggest problems?
nostrebored··on 1-Bit Bonsai Image 4B Image Generation for Local Devices
I don’t think you have ever actually lived in one.

“which plays on this general tendency of humans to prefer comfort over challenge, confirmation over rejection.”

This is completely and observably wrong. It reads like it would make sense, but the most ideologically open cultures I’ve seen are LIT.

nostrebored··on Launch HN: Expanse (YC P26) – Unlock Wasted GPU Capacity
this is true of most actual clouds/neoclouds. oversubscription and intelligent placement of workloads is already something they do. I’ve known a few people at AWS who have offset unbelievable amounts of cost by optimizing placement.
nostrebored··on 1-Bit Bonsai Image 4B Image Generation for Local Devices
I've lived in a low information trust society and this was not the outcome at all. People trusted their local cliques much more, and there were local minima (e.g. a mainstreamish political party with leaders that are actively and dangerously anti semitic), but in general people were way more willing to engage with ideas.

One implication was people were more social and talked about ideas more. Thought had not been outsourced to arbiters in the way that it was in the U.S. People with authority, knowledge, and close family members were definitely inputs into what people thought, but by-and-large people still came to their own conclusions.

You got to see the gradient of thought that people actually had about issues. People would say their insane ideas out loud. You could disagree with people and have them actually engage with your perception of reality.

It was strictly better in my opinion.

nostrebored··on 1-Bit Bonsai Image 4B Image Generation for Local Devices
I've lived in low information trust societies, and they were much intellectually healthier than high trust societies, at least in the white collar communities I was in.

Moving there, I was shocked at how "conspiratorial" everybody seemed about everything. Living there, I was shocked out how often they were right. But it didn't impact people's likelihood to do things. I think they are actually orthogonal in a way that is unintuitive.

nostrebored··on 1-Bit Bonsai Image 4B Image Generation for Local Devices
But this is exactly my point -- in a low information trust society things like this our noise. They work in high information trust societies because who would sell medical products that don't work? A doctor is leading Wamp®, surely a doctor would never lie to us for money?
nostrebored··on 1-Bit Bonsai Image 4B Image Generation for Local Devices
I am pretty excited. The factuality of important events has been distorted for most of history. Moving to a low information trust society is something that I think will be positive.
nostrebored··on Show HN: Open Envelope – an open schema for defining AI agent teams
Read some of your other stuff. I think we're on the same track, which is interesting! Everyone in our (admittedly SF centric) circles is trying to chase this down from the model path rather than building the consistent execution layer that we believe all of these solutions will need.

Wishing you luck with the project!

nostrebored··on Show HN: Open Envelope – an open schema for defining AI agent teams
I think this is kinda wrong.

Declarative approaches require validation to live at a synthesis layer, while an imperative approach that compiles down to declarative configs at runtime gives you the best of both worlds -- this is why anyone who does not need terraform cross compatibility will write things against CDK or Pulumi that has the same declarative schema wins with the niceness of testability and author-time typing.

Edit:

That said, it is shockingly close to the schema that we wound up with with a few ideas that I think are interesting.

reportsTo allows bottoms up orchestrator delegation

workspaces are interesting -- right now we have one bag of data with per-subagent data subscriptions, but this means that we frequently add input requirements to subagents that really should be more implicit

accessPolicy seems like a footgun to me -- i feel fairly convicted that tools should define their access scope and the only thing a subagent should know is the bag of tools available to it.

human approval seems redundant given we already have input requirements, and one can just be `email_approved` with a tool that emits the human approval request and `email_approved | email_not_approved` -- same feeling about `GateTypes` in general. If we are working on flat input-output requirements, then why do we need a specific GateType handler?

Trigger `any_approved | all_approved` is going to bite you if you move into plan solving. It is not rich enough to express XOR style relationships and I am willing to bet that v2 of your implementation splits TriggerRequirements where TriggerRequirements can be recursively applied to the type.

It seems like the tool definition is missing a lot of niceties that have been important for us -- for instance, at most once invocation. But we are working primarily over voice where there is a strong need to control execution for quality of service.

nostrebored··on MCP is dead?
this is bad. Anyone doing any cursory work with agents will realize how brittle <<just managing your own prompts>> can be. Adding an extra layer of indirection isn’t helpful, it’s a gigantic hindrance that gives you a moving eval target. Being an MCP developer means you have a moving target of model optimization. It is a win for nobody.

The tools we need to solve this problem exist and they are boring. Types, jsonschema, openapi, all of it is a better integration point than MCP.

nostrebored··on Green card seekers must leave U.S. to apply, Trump administration says
no, there are already runoff places that will brain drain the rest. there will be no "great repatriation". it's not just "The US" and "Home", there is an ease of immigration, quality of life, and success potential gradient.
nostrebored··on Python 3.15: features that didn't make the headlines
It’s completely the opposite. LLMs write awful python. Horrendous. They write pretty reasonable go and do it quite quickly.

We started in Python because of “the ecosystem”. It was a mistake. The amount of time we spend ripping out each dependency and pruning it to what we need is way higher than if we’d spent the month building out what we need. I miss compilers and LLMs will NOT generate config driven code or things that serde well. Everything has layers and layers of adapters by default and the domain model slowly erodes over time.

nostrebored··on Was my $48K GPU server worth it?
$6500 AUD can get you a good chunk of B200 time on any of the GPU neoclouds :)
← PreviousPage 4 of 34Next →