nl · HackerLangs

4 hari yang laludiscuss

I wouldn't be too fixated on the specific numbers in that post.

Anthropic was extremely capacity constrained at that point. They still are but not to that extent.

I'd note that OpenAI offers 24 hour caching. I'd be surprised if Anthropic hasn't optimised their caching for Claude code too.

SemiAnalysis recently posted that their actual Opus usage works out at $0.99 because of caching.

The principles remain though.

4 hari yang laludiscuss

It's not one prompt, but here is a parametric rod connector:

  Use SCAD and design a connector for square rods.
 
  The rods are 18.2 mm square. I want to connect two end-to-end.

  make if the bolt holes are created optional for each side - I want to set them separately. Make them M3.5 countersunk

  it's the +X or -X sides I want to turn the screw holes on or off.

  on each of the 4 sides of the connector add additional connectors at 90 degrees. Make each optional

etc etc

5 hari yang laludiscuss

I think that applies to military involvement abroad generally.

If you are dropping bombs on someone I'm unconvinced the use of AI will make them like you more or less.

5 hari yang laludiscuss

I've been doing a lot of 3D design in Codex GPT 5.5 (I found Opus 4.7 wasn't as good - haven't experimented much with 4.8 or Fable).

OpenSCAD is a parametric CAD programming language, and the models know it well.

The biggest challenge is communicating words like "inside" and "above" to the model - inevitably it's idea of which direction is which is often different.

I can't say I've done anything very hard, but for things like ESP32 cases, or parametric rod connectors it is great.

You can do things like "add snap connectors" and it'll do a great job.

5 hari yang laludiscuss

It's been very successful at frontier math tasks - a bunch of the Erdos questions have been solved by it - more than any other model.

https://www.erdosproblems.com/

5 hari yang laludiscuss

Their methodology isn't published.

Its widely accepted[1] that it runs the same query through the model in parallel and then has a model that either selects the best answer or synthesizes an answer from the multiple ones generated.

I believe most people think it runs 6 sub-models, but I think that is based on the pricing.

It's a pity that OpenAI doesn't publish details like this.

[1]eg https://news.ycombinator.com/item?id=48799977

5 hari yang laludiscuss

The source is the GPT 5.5 System Card:

> We generally treat GPT-5.5’s safety results as strong proxies for GPT-5.5 Pro, which is the same underlying model using a setting that makes use of parallel test time compute. As noted below, we separately evaluate GPT-5.5 Pro in certain cases because we judge that the setting could materially impact the relevant risks or appropriate safeguards posture.

https://deploymentsafety.openai.com/gpt-5-5/model-data-and-t...

There have been multiple podcasts with people from OpenAI which have confirmed this.

7 hari yang laludiscuss

I like this format:

"I love Lean because <abc>. I found it failed in <xyz> case because <123>. I created a thing <blah> which handles that like this: <ahhh>.

I'd love feedback! It's open source here: "

8 hari yang laludiscuss

Another Australian here too.

Yes, contacting your MP and senators can be very useful, including for federal stuff.

It's harder to actually get meetings with Federal members (they spend a lot of time in Canberra) but still worth trying.

Also it is very effective to vote for independent senators. You need to pay careful attention to make sure they aren't secretly insane but senators like David Pocock and Jacqui Lambie are very effective (Lambie seems crazy sometimes but she is surprisingly willing to change her mind on issues).

8 hari yang laludiscuss

No?

We do GDPR-compliant reporting by using differential privacy to provably remove PII.

8 hari yang laludiscuss

Related: https://news.ycombinator.com/item?id=48517377

It's too bad this has become political.

I do differential privacy work for GDPR compliance and it's an interesting technology.

8 hari yang laludiscuss

> variabilities to the chaotic circumstances of real world (“general”) problem solving. All forms of intelligence relate to the reduction of uncertainty.

Lossy compression is a form of generalization which handles this exact thing.

9 hari yang laludiscuss

You are going to need to explain yourself in more details.

> Compression is not intelligence

Just saying this doesn't make it so.

It's widely accepted that compression and intelligence have a close relationship. I think this summary of Marcus Hutter's work provides some background: https://www.antoinebuteau.com/lessons-from-marcus-hutter/

10 hari yang laludiscuss

> intelligence negotiates probability (allowing multiple divergent outcomes) while compression requires an idempotent symbolic translation.

What does this mean?

Lossy, non-deterministic compression is a thing. Does that meet the "allowing multiple divergent outcomes" criteria?

10 hari yang laludiscuss

Yeah, broadly agree. See my comment on the other story: https://news.ycombinator.com/item?id=48742711

10 hari yang laludiscuss

It isn't distillation that gave GLM 5.2 it's jump in performance.

To quote Pat Toulme:

There’s a big misconception about how GLM 5.2 was trained. Yes, they distilled Claude and GPT 5.5 — but distillation is not how they matched Opus quality. Distillation only fixed the cold start problem in RL.

RLing an agentic coding model isn’t rocket science. In simplified terms:

1. RL needs trajectories — rollouts where the model actually completed a task in some env

2. No successful trajectory on a task = zero gradient = you can’t RL it. This is the cold start problem

3. Distillation solves it. You seed your model with knowledge from a smarter one (Claude, GPT) on tasks it can’t do yet

4. Now it produces positive trajectories on those tasks

5. RL on those trajectories and hill climb agentic coding

6. At that point you no longer need to distill and can solely hill climb RL to better models

This is an interesting curve. I’d argue it’s harder to get to Opus 4.8 from scratch than to go from Opus 4.8 → Fable/Mythos tier.

GLM 5.2 is already producing positive trajectories, so they have plenty to RL on — they’ll keep climbing to Mythos quality without distilling any further. They no longer need American models.

https://x.com/PatrickToulme/status/2069211575437627743

Not exactly sure what the finish line in "the race to superintelligence" looks like and even moreso it's unclear why you think being there first is a critical benefit.

10 hari yang laludiscuss

Opus is still significantly better than open weight models.

GLM 5.2 comes close on agentic tasks, but doesn't code as well.

Kimi 2.6 and Deepseek v4 Pro write great code but lose track when doing agentic workflows. They were better than Sonnet 4.6 but not as good as Opus. I haven't compared them to Sonnet 5 yet.

10 hari yang laludiscuss

Opus isn't nothing!

I think the 5x subscription is here to stay - I'd bet they make money on that from lots of people not using it.

The 20x is already unavailable in Teams plans.

10 hari yang laludiscuss

This post has been marked as a dupe, but it provides a lot more details than the other announcements of Fable's re-enablement provide:

> The export control directive on June 12 came after the government became aware of a report in which Amazon researchers had found a method of bypassing Fable 5’s safeguards: prompting it so that it identified a number of software vulnerabilities. In one case, the model produced code demonstrating how the relevant vulnerability could be exploited. Over the past two weeks, we have worked closely with the government and other partners, including Amazon, to review the report and evidence.

> Our testing confirmed that many less capable models—including Claude Opus 4.8, GPT-5.5, and Kimi K2.7—could identify the same vulnerabilities as Fable 5 did in the report. When it came to the demonstration of how to exploit the single vulnerability, every model we tested could produce the same demonstration as Fable 5 (including Claude Haiku 4.5, Sonnet 4.6, Opus 4.6, Opus 4.7, Opus 4.8, GPT-5.4, GPT-5.5, and Kimi K2.7).

This indicates three things:

1) WTF was Amazon thinking? Didn't their researches try the same thing in other models too before telling the CEO to tell the government it was dangerous (!?)

2) Anthropic - in particular Dario - really needs to learn government relations better. Most of the problems Anthropic has had with the government seem to stem from Dario's attitude rather than actual facts. (Eg, the DoD debacle seems to have ended up with OpenAI signing almost the same contract Anthropic already had, just worded differently)

3) The administration decision making is just wacky. In a normal administration they'd have actual policy documents you could look at to understand under what circumstances they think models have a problem. With this they just seem to make it up as they go, and the tools they use make no sense at all. If it is dangerous for cyber security reasons why would export controls make sense to use?

10 hari yang laludiscuss

> In contrast, I've never seen a large model turn bad instructions (instructions that would cause a human to think before starting) into a result I liked

I think the distinction is here.

I expect my agent to build from product level descriptions. This might include specific special cases that I call out, but will rarely highlight existing special cases or edge cases - they already exist in the code, and I'd expect a programmer to make sure that behavior continues to work.

If a feature hits lots of these edge cases, the weaker model that is reading the code (aka Haiku) won't understand their significance, and will report back to the planning model incomplete or incorrect information.

The planning model (Opus - which hasn't actually seen the code remember!) will build a plan that is incorrect or incomplete and delegate coding to the mid level model (Sonnet) which will do it's best to make things work, without understanding the overall picture.

This is how you end up with slop - for example Sonnet reimplements things that already exist because it found one of the edge cases, but Opus had never known about it because Haiku didn't understand it.

It's possible that the new "agent teams" feature in Claude code can help with this. That keeps each agent alive with its context so they can ask each other things, but I haven't tried that enough to be sure - let alone with the specific model mix like this.

In your case, you are giving the Sonnet model specific instructions for what to implement mindlessly. I'd expect that to work well!

But that's not the same as the agentic workflow many other are using.

nl

Submissions

AI Value Capture

Noise infusion banned from statistical products published by Census Bureau

The household battery revolution that could change energy bills and the world

Anthropic says it's about to have its first profitable quarter

OpenAI Stargate: where the US sites stand

We Tested DeepSeek V4 Pro and Flash Against Claude Opus 4.7 and Kimi K2.6

US/China talks to make sure non-state actors don't get a hold of these AI models

Show HN: An Interactive Text to SQL Agent Benchmark

The Pentagon's UFO Psyop

Stress testing Claude's language skills

Chat with Llamma 8B at 16,000 TPS

Llamma 3.1 8B in hardware, 16,000 TPS

DeepMind Aletheia [pdf]

Accelerating Scientific Research with Gemini: Case Studies and Common Techniques

Agentic coding is accelerating app releases

Four Ingredients for Successful Retrofitting

In Defense of Data Centers

Erdos 281 solved with ChatGPT 5.2 Pro

Microsoft's spending on Anthropic AI on track to reach $500M

Tim Dettmers: A Personal Guide to Automating Your Own Work

comments