Hacker News

> Some people would say that you don't need this at all with the frontier LLMs

Hit the nail on the head. Development methodologies pivoting at the whim of frontier models have been one of the things that have been worrying me as of late.

> Feels better to me in the sense that it doesn't try to be too clever. When I ask it to do things it actually generates code that a developer would write. That still lets me build software in layers like I do mentally. And it keeps you in control of where abstractions are.

yes!So it's not just me. In terms of how intelligent they want to be OpenAI is at the top, then Claude. And actually when I was playing with Sol it meant that many tasks it would just try to do something else because it thought I don't know what I want. Even trying to nudge it back in the right direction wouldn't work.

I guess the models just have the cultures of their creatures.

> it doesn't try to be too clever

DS is miles better than GLM in this regard. I wanted to add that as PC is planning to go to GLM next.

What harness are you using?

Z.ai does offer access to DS AFAIK, so I guess that's what they mean? Maybe the feeling is that it provides more value that DS's own platform since you also get GLM, not sure.

maybe the point was become of the openness, they will be using z.ai?

I'd rather my data go to China and directly to labs that release open weight models (DeepSeek is MIT licensed still); instead of closed labs and the NSA/CIA; who has far better ability to hurt me than China does.

Without getting too much into politics; over the past several decades, one country is going around starting wars, regime changes and other interventionism, or initiating economic warfare (Huawei sanctions because they were getting too competitive to US companies; GPU restrictions; tariffs; etc), and that's not one administration: it's US policy over the past few decades.

To add to this, Fireworks (and many competitors, e.g. enumerated here[0]) offer zero-data retention inference, to anyone. DeepSeek cloud does not. It's part of their terms that anything you send them is fair game for them to train models on.

[0] https://openrouter.ai/deepseek/deepseek-v4.1-flash

(The badges next to companies indicate their privacy terms, i.e. green/yellow/red)

I figure (hope) buying from DeepSeek directly benefits them the most so they can release even more improvements and open-weight models.

Plus only one country has threatened to annex mine and it wasn't China.

If one is not American, why prefer American? Chinese fine by me, I trust the US providers just as little.

There are multiple others in EU, and in US. Cheaper, more expensive, slower, faster. So much choice.

Completely unrelated, but it's so odd to see upside down '8's in the first column of the base tables on their pricing page.

The link points to OpenRouter with multiple providers.

> they frequently generate CoT explanations rationalizing those answers

I can't tell if this is a really human thing to do, or if it's just a sycophantic thing to do when your master likes a good story.

Fair, but you can see the logprobs, right? That alone is pretty valuable.

This is like one of those folk theories that have some truth but in practice it is useless.

I love COT and gives me a good idea about the thought process. Sure it doesn’t work always

In practice, it does help to immediately be able to tell if you missed something or wrote something incorrect in your prompt, compared to if you didn't have access to the CoT. If you see that it goes off wondering about stuff or thinking about stuff that you know are already solved/known, you cancel the execution, edit the prompt and redo, then observe the CoT again, until you've resolved any "misdirections".

Both OpenAI and Anthropic hides so much of that, that you don't really know if it misunderstood something or you got something wrong, until all the turns are complete or you directly see tool calls that implicitly tells you something is wrong.

You don't need to read any papers to see this effect concretely when developing together with LLMs.

[dead]

I've used the zai subscription, opencode go und cline-pass in the last few months and would recommend the latter. Zai quickly drains your quota as soon as the context window is at more than 10-20% and opencode was a bit unreliable at times.

I wonder what quantization opencode uses for DS models.

How would you go about that?

Depends on the question. Most models won't tell you how to make weapons because they're trained not to, but the political stuff is usually blocked on the API level

to add to the corrent sibling answer, the unhinging is called abliteration by the wider community but it's hard to find someone who will serve that over API.

It will not. Ask DeepSeek about Tiananmen Square. It will happily answer

GLM refuses to write exploits. It's not open season with anything you want.

I mean... I ported some of my childhood Finnish shareware games for Linux and Rust with DeepSeek V4.1 Flash. It took about six hours of spelunking and finally I was pretty happy with the results! I did it twice. Tokens for both of the games were around 3 euros.

What about mimo v2.6 pro?

Muse is fantastic for open source work. At least 1.3 was, for its price. Wonder makes less economic sense.

I don't see this kind of behavior myself, I wonder if it's me not pushing the boundaries enough or me explicitly or implicitly framing the prompt in a way that makes it less likely for the agent to intervene.

Z.ai (well technically alibaba) used to support both GLM and DS models in their API and coding plans.

It happens very frequently if you happen to work on anything low level, e.g. a C code base, a database code base, etc. that deals with memory, and these have nothing to do with cyber security. It is extremely annoying and makes these models unusable.

E.g., your prompt for simple auth works (e.g., decode JWT), but your sessions stop because you are flagged for cybersecurity risk.

The Fable is the worst offender, and it is not purely Anthropic's fault.

It thinks a bit more, but if the cache reads are free or almost free, and it blasts through 10x faster than GPT Sol/Terra/Luna, it doesn't really matter. Just don't use the max thinking for implementation, go with high or low.

I think, as long as you don't set thinking to max and use a fitting system prompt, it's not a big issue.

Yes, and this version explicitly addresses that

4.1 Flash on medium is a solid builder agent

LLMs don't think.

Note that opencode routes deepseek requests to China.

I can think of plenty of tasks that even the closed frontier models fail miserably at. From my perspective we have a long way to go for these models to get better.

Through fear. Specifically making the American public so scared of AI that they hand over regulatory keys to the only “experts” that can keep them safe. This is a centuries old political strategy used by some of the worst regimes including apartheid South Africa where I grew up. It’s horrifically effective.

You are taking their words literally, rather than looking at the side effects of any effort to "pace the frontier".

Dario has not (and will not) say he wants to ban Chinese models outright. But:

1. He's made many posts about how he thinks "dangerous" models are bad (https://www.anthropic.com/news/position-open-weights-models)

2. He writes posts proving Chinese open weight models are "dangerous" (https://insightsintegration.com/anthropic-warns-of-advanced-...)

3. He talks about about how he wants stronger govt regulation of AI (https://www-cdn.abcnews.com/Business/exclusive-anthropic-ceo...), and explicitly mentions he does not trust China at all and doesn't want them to have Mythos.

Do the math: 1 + 2 + 3 = he wants to regulate/ban the Chinese open weights.

Of course, Anthropic and OpenAI will then write posts about how their own models are dangerous - but that's okay, they're the good guys.

Go read Dario's screed on GLM, or his other one one on how your toaster could be part of a distributed open model that self replicates. Every second blog post is some doom scare about the dire consequences of open models. Go look at the new ToU [0] which are now pages and pages of things you can't do with Anthropic models. Check how they will ban your account if you are persistently mean to Claude now.

He's completely transparently campaigning for regulation. It must be very frustrating to him that the Trump admin is completely non-responsive. He probably thought Anthropic would have the whole market stitched up and water tight by now.

[0] https://www.anthropic.com/legal/aup

Literally zero, but that never stopped a good conspiracy theory, especially when half of HN hates the big labs because of fear for their own economic future.

OpenAI and Anthropic aren’t even slightly worried about DeepSeek.

We have to keep remembering this for some reason.

New model comes out and someone says, "it can (or can't) draw ducks on bikes" or some esoteric thing. But it's not enough statistical power to claim anything other than a personal anecdote.

I got a flat tire in downtown today. Therefore, driving in downtown causes flat tires.

Honest question about the non-deterministic nature of llms. Every programming language provides some way to control the random seed. Is this not true in GPU programming land?

How do you get dirty changes into a public repo online?

Especially when $10-$20 on OpenRouter with e.g. DeepSeek (or any other inference provider, or their official API) will get you very, very far.

That may be the case, but simply running the same code review twice with the same model may give very different results as well. Impossible to know with the information we have.

If none of the findings overlapped then LLMs are either

- For some reason choosing only a subset of issues.

- Different LLMs actually learn different things.

I suspect the former since they do this for me and it's extremely annoying sometimes

Very commonly used acronyms in business/SWE contexts.

The lower the number the higher the priority, e.g. if production is broken someone might say "This is p -1" (higher than your highest priority).

P0 is "do this immediately", P1 is very important, P2/P3 are like "nice to haves." P100 would mean something like "we will never do this"

priority 1,2 etc, more like severity of the cases

It’s a very common abbreviation; i had no issue understanding GPs post tbh

I tried it out today, was the new one: 1017.

But with that version number, it's not a "minor" update! It's a patchlevel update! -Engineer

Strictly speaking, it just shows they aren't marketers. Bad copy doesn't tell us anything about their engineering skill.

So they meant a "small" update - "conservative", "inexpensive"...

An example when using a less frequent term actually reduces specificity.

Remember when we were dumbfounded why our service is slow despite having implemented quite robust caching mechanism. Turns out simply the flag that enables it was hardcoded to false. So it was minor update (changed flag to true), but provided substantial speed up.

Source?

Then I guess all that it changes is that my proposed setup gets even better?

Huh? This is a new checkpoint. It has identical architecture, only different weights, and most likely just more post-training.

Much like Qwen 3.8 Next; which is qwen4-exp in a lot of the code/patches, there's usually a tick-tock cycle with LLMs. That said, it's hard to know for closed frontier models.

Without any context for what you're working on, I'd have to assume the reason you don't hit the $20 subscription limit is that you're either not using it very often, or you're working on some relatively basic CRUD app that doesn't require it to think very deeply.

I tear through my $200 subscription because I often have 3-4 agents running simultaneously, exploring performance optimizations, building features, monitoring production rollouts, or investigating production problems/customer issues in parallel.

The downside is that I do spend most of my time reading the agent's output these days.

Its easy when you spawn a couple of subagents in a homelab with a rack with 5+ servers.

This is how one of my claude code session looks like right now, one prompt from yesterday:

  ◯ lean-tracks  ▰▰▰▰▰▰▰▰▰▰▰▱▱▱▱▱▱▱▱▱  7/12 · 18h35m · ↓ 3.1m tokens
  ◯ lean-tracks  ▰▰▰▰▰▰▰▰▰▰▰▰▱▱▱▱▱▱▱▱  6/10 · 17h47m · ↓ 2.6m tokens
  ◯ lean-tracks  ▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱  0/1 · 4h49m · ↓ 275.9k tokens
  ◯ lean-tracks  ▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱  0/1 · 4h30m · ↓ 395.0k tokens
 ◯ lean-tracks  ▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱▱  0/1 · 3h50m · ↓ 490.3k tokens

I swear I have the same question in my head always when I read such statements! I have never once reached 90% that includes everyday usage across multiple repos, artefact generation, device debugging, long running harness work!

I am not saying OP is not telling the truth. I just really want to understand!

Not OP, but I used to run out in a few days, for me it was a matter of suboptimal code review practices with many subagents looking at different aspects, often expanded to look at more than the diff.

skill issues, brute forcing with AI.

not reviewing its ongoing work and spawning subagents that end up looping each other

Loading tons of MCPs and never compacting

[dead]

[dead]

I can drain a $200 OAI plan in 2 hours flat working on my 10 projects simultaneously.

Maybe you're not being ambitious enough with these things? They are at their most fun when you spec out sprawling architecture documents and watch 1+8 astras buzzsaw through it.

I'm only at 260B for the year, though. I'm sure someone who has used more will be along shortly to chime in

E: The other subthread talking about tests is correct. Expensive models should not waste a single second waiting on tests or validation. If you cannot assign a single cheap model to be the test bitch, then just tell them to frontload everything and clean up later as a single wave.

By not using existing tools and one shooting everything through the API - I.e. instead of decompile via Ghidra they will try to get source via capstone genersted assembly.

It is like putting a jet engine on a shipping container and wondering why it needs to be constantly kept in afterburner to at least scrape on the road while smarter users put wheels on the shipping container.

> It's not any more complicated than that.

It is more complicated than that when the valuations and business plans are contingent on those companies AI having complete dominance at every level.

If companies only use frontier models for occasional difficult tasks, not only does that affect volume, it basically elminates stickiness and incentivizes companies to instead of using their products continuosly, find ways to move from the expensive stuff to the cheaper stuff.

Yes but that’s a tiny portion of AI use cases. Maybe 1-2% max. The big labs need a lot more than that to survive.

You’re assuming other model creators can’t build these mega models. Evidence is they’ve simply decided it’s a fools game to do so and they’re focused on where the use cases are.

Fable is behind Opus (which is also much cheaper) at this point in benchmarking.

That trajectory appears to be changing. Virtually every company I see is shifting more and more to open weight models. Anthropic’s slowing growth rate is already spooking investors and that shift is just getting started.

Also they can work in completely new domains, like Astra and Opus 5.5 starting to be able to use something like Blender. Currently they still often do horrible topology and basic UV mistakes, and the animations are really bad, but who knows where they’ll be in 5 years.

Very cool!

Is your laptop something like framework which has open source firmware?

I(and many other people) have a similar issue with usb-c charging not working consistently. I've been meaning to try one of these new smart models to help me debug the issue.

How are these high-speed providers close to matching the API rates from deepseek?

Where does one even get decent ZDR provider? I'm wary of sending sensitive data but not at a scale where I would build my own machine.

I use Opus for planning, through API on medium thinking, it doesn't cost anything significant and the rest can be happily done with deepseek or glm flash. Using sol 6.1 for review loops at the end. All running on pi.

A very complex task tends to cost like 10$ max. It's comparable to a subscription with enormous flexibility.

It's a steal.

Honestly, I don't know what people see in Sol 6.1. Opus 5.5 is so much more inteligent, as in proactive, avoiding stupid mistakes, where it contradicts itself, ignores the context. Sol 6.1 has a gold-fish memory, it's as annoying for me to work with as Opus 5 was.

Yeah performance at a given compute budget will likely continue to improve.

We are running Qwen 3.8 27B FP8 with 4x 5060 16gb at the office for a small development team. It is doing pretty well for implementing backend and frontend tasks, and is already useful. In 6 and 12 months from how I expect it to get even better. Some planning is still done with a stronger model, but even that is maybe not strictly necessary.

No, this is not a small model. I believe 4x Sparks is possible but 4.1 is pretty big for a flash model.

MiMo 2.6 Flash is the most capable model you can currently run on 2x Sparks without resorting to a quantised version of something.

SSD weights streaming. The actual weights is 552B, then engram.

By “quantitizing” it

here you go! Results are trickling in https://fenderog.com/art-crit/page.html?p=surfing-pelican

Faithful to the original theme. Apropos. +1 on adding the prompt as long as it’s affordable and easy enough Simon, and not too distracting/noisy/time consuming all considered!

Some of the most recent Claudes and GPTs have done that too, and Gemini has done it for a while.

What did I miss?

[flagged]

It costs money to run serious tests. We're way past the time when you could get an idea of the new model in a couple of tricky prompts. Now I need to compare a whole resulting project to get a clear view. If someone else proves the benchmarks don't hold up in reality, then it saves us further tests.

Tavily one is free, too. Easy to use cli with a skill.

Pretty funny to say "us peasants aren't concerned with ... cost/token", as if most people ("peasants") won't find it absolutely insane to be paying $200/mo for a subscription.

Caveat: I get a budget of free ai tokens at work, as I work for an open source AI inference provider.

My goal is to reverse engineer an entire gamecube game -- I'm pushing an existing community decomp that was stalled to the finish line, solo, with lots of AI help.

This work was directly inspired by https://decomp.dev/ -- which has just shut down a few days ago because of, ironically... people like me who enjoy decompiling things with AI, so I'm still weighing / grappling with the ethics of how to release my work, it's murky waters at best.

I use Opus 5.5 as the orchestrator, as well pi harness because it's somewhat lightweight. But for the really hard functions I don't use deepseek I use opus managing other opus agents.

I'm working to release something soon as proof that I can do a 100% decompilation job, but the main idea is that when decompiling a gamecube game your sources of truths are individual functions or files, so you can build off that primitive.

Each ai gets a worktree, and a shared giant context file that's basically a record of every experiment that has happened. They're assigned tasks to essentially do as much as they possible can for each individual funciton. After about an hour claude uses a script to merge the work that's actually correct (verified by tooling that checks if the c goes to asm) and then there are usually 3 rounds of reviews by independent ais, because there will always be some slight mistakes. The techniques used as well as the mistakes the models make are recorded in the giant ledger. There's a bunch of random one off scripts such as stuff to clean up, and tools to make sure the agents don't accidentally google stuff / delete stuff / use too much space.

Claude steers the ship via ssh on my strong linux tower at home so it keeps my laptop nice and cool and at this point I just check in from time to time, I am basically a manager, and Claude has it's own workers.

Last point -- I don't know any c besides the basics, I've only been "coding" python / ts for a couple of years, but even then with the extreme strength of models the future is managing agentic swarms, so I'm leaning in as much as I can. I really see this as the future of large scale projects. For some people it's a mockery of decompilation, but, my goal is the finish line, not the journey.

See here https://news.ycombinator.com/item?id=50135410

Yeah, I had a somewhat odd conversation with a Chinese acquaintance last year when she got to talking about how America sees China. Her point was that, "How am I not free? I can do whatever I want, go where I want, buy what I want. I can talk to you and say whatever I want; nobody is here listening to me."

The fact that she can't call Xi "Winnie the Pooh" on Douyin wasn't something that occurred to her, or maybe that she was aware of at all. I didn't ask. I don't like to get into these discussions when I'm in China.

That economic freedom is a fantasy fabricated so that the king can rob people fortune created by the themselves in the name of greater good.

The go-to is a local model.

Both the free (seems limited to ~200k context) and metered variants in opencodego show the release date as 1788998400000 → 2026-09-10, which implies they're not the newer checkpoint

The only way I'm aware of is to check their twitter feed. They moved to Deepseek v4.1 flash in September

[dead]

These files came to be after I've seen Deepseek create them for itself to track progress.

Then I turned what Deepseek did into an upfront policy.

This seems to work quite well generally: watch the agent as it works, and think what could be provided to make its life as easy as possible

The idea behind using Luna was cost-optimization.

I wrote a Pi extension called "model-escalation" which lets the model request an automatic switch to a smarter model for the next turn if it thinks the problem it is struggling with is too complex or if the user explicitly asks it to do so.

With this extension in hand, I could split the work between Luna and Sol, resulting in a ~95%-5% split. I don't have to switch to Sol when review time comes, Luna does it automatically. Sol tells its remarks, then Luna gets the control back and continues the back&forth (if there are still complaints).

Deepseek servers went unavailable for 5 hours. They compensated that by rewarding the previous week usage as credit though. Pretty nice.

Ai/tech stock crash

OpenAI's models -- and even Google's Argon -- have particular niches where they shine. For GPT-6/6.1, it's 3D modeling; Astra is hands-down the best at making .step files, Blender files, etc. For Argon, it's natural language writing and everything that hooks up to Google, e.g. search and maps-based itinerary planning.

Anthropic and DeepSeek (+ GLM) are definitely solving coding, though!

They're all quite different tools.

I believe these are the last sets of super-high-end models available to public (open or closed).. these companies probably will now start to withhold top-of-the-line models for themselves, governments, and their super-high-paying partners only.

idk, sol 6.1 is both cheap and capable. i've been using it exclusively, i don't need to care about usage, and it's really smart.

yes maybe opus 5.5 is stronger and flash is cheaper, but 6.1 strikes a good balance imo.