HNHacker News
TopNewBestAskShowJobs

nostrebored

2,711 karma · joined March 20, 2015

Opinions discussed here are my own.
submissionscomments
nostrebored··on What is nueralese and why is it bad
Lesswrong is just so comically collectively incorrect.

This makes no sense. There is no guarantee that reasoning aligns with an outcome. Reasoning is effectively saying with more compute and ability to change attention on the fly by altering context, we can come to better answers. Reasoning often has a nice property that the English intent is aligned with what the model wants to do. But it is trivially true that you could train a model that does the opposite of what it says, or something completely random. The interpretability is incidental.

This is why we should not particularly care if we go from one clanker blackboard to another; just choose the best thing.

nostrebored··on Qwen 3.8 27B available on Cerebras at 1500 tokens/s
Celeris reaches ~50% of the speed on commodity hardware. celeris-magnus-1 is based on qwen3.8-27b. Maybe we will get there without custom silicon!
nostrebored··on Solving the Jane Street reverse engineering challenge
z3 is also just so thoroughly optimized that even if your formulation of the constraints is inefficient it is faster. it is a great library that lets you solve pretty complicated DP problems with a few dozen lines of code.
nostrebored··on Qwen 3.8 27B available on Cerebras at 1500 tokens/s
in my last tests with cerebras for coding tasks, most large tasks or anything greenfield would hit token limits. note that smaller models and the gpt-oss-120b style models they used to run are very prone to overthinking, so individual turns may be 3-10k tokens of just thinking + input + output.

i don't think it's quite apples-to-apples to compare to a frontier model or even a k3. the odds of success (file compiles? read the right context?) are lower and thinking is longer.

nostrebored··on Qwen 3.8 27B available on Cerebras at 1500 tokens/s
iirc input (uncached) goes towards the limit as well
nostrebored··on Qwen 3.8 27B available on Cerebras at 1500 tokens/s
150k TPM limit on public endpoint means that it's likely unusable for many coding tasks. When we've tried Cerebras in the past, our problem has always been rates. We'd love to not deal with dedicated and to have access to a more flexible rate pool.

Even trying it out, it seems like our account has gotten moved to some limbo where we can no longer add billing information.

``` Billing access restricted Self-serve billing is not available on Enterprise accounts. Please contact your team for further questions. ```

We have no team (they removed themself from our slack channel after we talked about rate limits). Perplexingly, none of this even shows up in the request, which gives:

``` {"message":"Model does not exist or you do not have access to it.","type":"not_found_error","param":"model","code":"model_not_found"} ```

When the error is really about billing.

I always want to like Cerebras, but I get the vibe that as a tokens in tokens out consumer you are not valued at all.

nostrebored··on True Rate of Unemployment
As someone who grew up below the poverty line for much of my life — 2x the poverty line seemed like wealth to me. There are absolutely people getting by on “feels like” 50k a year.

There are not really frills, but in terms of “have shelter and food” it can work. Not everywhere is SF.

nostrebored··on Stripe said to abandon $50B pursuit of PayPal
Maybe we’re doing different things or in different regions, but I rarely see pay with PayPal now.
nostrebored··on U.S. State Department pauses immigrant visa applications
People are not as capable as you think they are and software engineering (well) is actually difficult. You cannot just push more people in. People are not fungible.
nostrebored··on AWS Acquires DuckLabs
Many Amazon teams have had high expectations for work. The stories I’ve heard have largely been around how AI is used, slipping CR standards, and what is funded.
nostrebored··on U.S. State Department pauses immigrant visa applications
1/ Training is only effective if you can fill the function of the job and learn from it. Many software engineering jobs people claim are stolen by H1B hires are filled by human beings more likely to meet the bar. Many American candidates just don’t. Filling a software role is hard.

2/ Demand for software is not static. Good software drives demand for more engineers in many cases, intra or extra-company. There is probably a non zero rate of wage suppression per marginal h1b currently. But an American economy with no h1b employees could feasibly have a much lower average sde salary. It is not an Econ 101 problem.

nostrebored··on U.S. State Department pauses immigrant visa applications
Having worked at a few large companies, this was the case at all of them. I do think there is a messy middle of companies built on vibes.
nostrebored··on AWS Acquires DuckLabs
AWS in large pockets was not a mess 10 years ago.

The things I hear from recent AWS departures are very concerning.

nostrebored··on AWS Acquires DuckLabs
Some of the largest, highest TPS, highest throughput systems in the world use DDB.

The constraints are what let this happen. Unconstraining it might make a better generalist product but part of what you're opting into with DDB is the dumb "put an item in get an item out semantics" and the other side is knowing that it will still work if that volume increases dramatically.

nostrebored··on Oceans hit highest temperature on record
There’s something odd about the concentration of true believers in ai xrisk and climate change catastrophism. They are both technopessimist in a way that is seen as progressive. Stopping progress in AI would make people’s lives immeasurably worse, while stopping fossil fuel consumption would kill billions of people. But it seems like a particular kind of risk aversion that is blind to upside.

I’m curious why this calculus makes sense to you?

I’ve been friends with eco-activists who are very against anthropocentric ideology. Their perspective was that the deaths of billions of people was worth it. But this brand of conservatism is one that I’d really like to understand better

nostrebored··on New MCP Roadmap
What do you mean by this? It certainly sounds technical but it seems to not mean anything.

MCP has not smoothed over legacy cruft, and it is generally bad at exactly what you’re describing (many unintentionally coupled APIs with unintentional side effects). These require near deterministic trajectories and you’d be better off creating a consumer with a series of well known good patterns with useful results.

If you take it a step further you may allow for a common language and keyspace of these well known results and employ dynamic solvers that are entirely agnostic. LLMs have made creating these much easier!

nostrebored··on New MCP Roadmap
Enterprises have been managing thousands of http endpoints for decades now. It is not easier. It would have been easier to have something swaggeresque that lives at the openapi spec layer but that’s not cool and AI.

What is not hard to understand is that EVERY MCP UPDATE is almost certainly a breaking change. The versioning story is not as mature. The models using it are different.

It is an unuseful fiction that by storing a blob of instructions next to a remote endpoint that things have been made easier.

nostrebored··on New MCP Roadmap
Your stateful remote execution needs a stateful tool middleware for your stateful agent, just to make sure that it’s impossible to test or evaluate
nostrebored··on Feature Request: Support AGENTS.md
rg CLAUDE.md > refactor_hits.txt

claude -p "@refactor_hits.txt support either CLAUDE.md or AGENTS.md in each location. if they both exist, prefer CLAUDE.md and do not concatenate"

nostrebored··on Claude Code May–August 2026 weekly limits promotion
prediction: extension of promotions because of the progressive move to codex.
nostrebored··on Don't classify, hallucinate
The idea of distance thresholding models that are trained to satisfy an ordering constraint is a bit strange. The reason it's hard is that there isn't a threshold!

You can slice and dice it a ton of different ways, but the significance of groups is incidental.

It's a good starting point, but having done this a few times for a few companies it always seems like it needs substantial human review.

nostrebored··on Retire the Abstractions
Professors have no reputation for writing good code. Academic code is fairly notorious for being under abstracted, hard to follow, messy, etc.

If a passerby told me this story I’d have the same guess though!

nostrebored··on Retire the Abstractions
why?
nostrebored··on Retire the Abstractions
it sounds like a bad abstraction. ime llms are really bad at reasoning through efficiency and utilization problems. they fail at scheduling and bin-packing for the same reason. it's actually difficult, multivariate, and not great.

i would be shocked if a well designed framework were less efficient than "bespoke" clankercode for low-level work

nostrebored··on Amazon backs power plant that may become top source of US climate pollution
It was sarcasm. It's also not particularly insightful. "AI data centers" get immense scrutiny right now. Astoria has a similar generation capacity in a much denser (and human populated!) area.
nostrebored··on Ten advances in mathematics and theoretical computer science
"Small" was a poor choice of words here, "low compute budget" is more what I'm getting at.

In voice interactions, ttfat is actually relatively important. If you look at models with a <1s ttfat you eliminate almost every reasoning model, less some of the diffusion models and more obscure ddtree/dflash like speculative decoding implementations.

nostrebored··on Ten advances in mathematics and theoretical computer science
For customer support I don't think models have gotten better since gpt-4.1. The class of small models, with limited to no reasoning, that need to handle a complex issue with a touch of empathy, has not improved much.

I think most are actually worth, as agentic harnesses seem to optimize for solving poorly described problems rather than following complex procedures as written. In other words, instruction following maximizing models seem to make worse free-form agents, but they're really all that some domains need.

nostrebored··on Show HN: Mu – Tools for Agents
but... why?

you almost never want these tools in any given session, and creating any of them can be done simply with Claude. looking at the PRs, it seems like that is exactly what's happening. i guess this is my generic problem with MCP releases, who are they for and why?

nostrebored··on Pacing the frontier
we're always in uncharted territory. that's why most HN commenters think that this is hyperbole.
nostrebored··on Qwen 3.8
It’s in the weights. Context needs to be attended to to create a response, and the weights dictate what response is decoded. If you include retrieved context that has an American perspective, I imagine the think trace has some reconciliation about how they must be incorrect.
← PreviousPage 3 of 34Next →