fxtentacle

At first, I was feeling kind of shocked that AI would do a useful and good reverse engineering and produce code that can be extended later, because that totally not matches my experience working with frontier AI models. And, indeed, they started with a working cross-platform GL-accelerated, multiplayer-capable and moddable codebase. And then used AI to remaster it. So the AI contribution seems to be mostly upscaling the old pixel art.

"OpenOMF (MIT): the open-source reimplementation whose reverse engineering this port follows."

https://openomf.org/

I’m hoping that someone will tackle Garmin entirely by releasing open source replacement devices ..

... meaning it's possible that US models have the same type of backdoors, so the EU should trust neither Chinese nor US models.

It's been 12 years, but I once wrote a blog post on how my commercial spatial audio effect works: https://hajo.me/blog/2014/12/28/how-surround-sound-for-headp...

BTW, the resulting plugin was used during mastering of George Lucas' Red Tails, so I guess sound quality was fine ;)

My own Mac app has long since been discontinued (because Apple locked me out of the KExt permission), but I sold the tech and it lives on.

I keep wondering if I should build hardware headphones with the spatial audio inside, so it looks like Dolby Atmos to the TV but then does virtual surround sound on your headphones, but hardware has low-ish margins so I never found anyone willing to fund it, even though I already have a working prototype. Oh and the audio engineer who invented it has since been Grammy-nominated twice. And, well, Hollywood likes it. But still a tough sell towards investors.

"What is this superpower, and why is nobody else using it?"

The "Undo" time-travelling debugger for C/C++ has been around for 20+ years. So I guess the answer is simply that most people chase shiny new things instead of using established, reliable, and, hence, boring tools.

Fully agree.

Since even Fable 5.1 cannot reliably remember what it wrote into "memory" files before the last compaction, why would we trust it to not forget about some side-effect when handling assembly that is so verbose it'll surely exceed the context and, hence, necessitate compaction.

I'd say you're fully off the mark here. That page contains serious architectural insight on how to make a "designed" car feeling work within a "physically correct" collision engine. Getting the balance right is super hard and he explains how to do it. I'm confident no AI could do that architecture, so this is most likely hours of hard work with a thin layer of LLM gloss on top.

Hi there, I have some experience with reinforcement learning, PPO, RLVR, and I trained both video game AIs and sim2real robot policies.

In my experience, videos from professionals are not enough to train a robot, because you lack all/most of the edge and failure cases. Let me illustrate with pilots: The average Airbus can land itself via autopilot. Autopilot will fly it to the correct location. You confirm the frequencies, it connects to the airport's guidance system and lands itself. Super boring. Videos of that won't help anyone become a proper pilot. The reason pilots train for years is to handle the 1 in a 1000 case where the autopilot doesn't work. And you suddenly have to take over in the middle of a messy approach.

Translated back to your plumbers, the real training is when the robot is carrying a 100kg metal pipe and then the owner's kids come running and pass through under it. Every professional will give their best to immediately reduce efficiency by making things stationary and SAFE. That's the training data that is truly valuable.

BTW, I am considering bootstrapping a startup that uses video game engines to generate training data to make robot policies anticipate physical effects and counter-steer before they happen. Like putting down a heavy steel pipe before it drops due to gravity. With 2 bugfixes merged into the Unreal Engine, I have some experience there ;) So I might soon be in the business of selling custom training data.

Feel free to email me at hn at hajo me

... brought to you by storyteller.ai, which now has a defunct hero video, but apparently still 10 mio users (at $12+ monthly). I very much applaud them for pouring their (est.) $1.4B ARR into Open Source.

Labs typically pay $2k for each [prompt+scorer] docker image. So this is why frontier labs need so much cash and manpower and they see it as their moat.

But some of the resellers have freebies, like:

https://app.primeintellect.ai/dashboard/environments?ex_sort...

https://github.com/sierra-research/tau2-bench

Deterministic feedback is precisely how frontier models are trained. It’s called RLVR. You let the agent run on a problem and then calculate a deterministic score of how well it did. Repeat 1000x times and you can “brute force” a good solution. (Which includes all thinking traces and you add it to your training data.) And then a Chinese model can copy your advance for 1000x less compute. Which is why US labs call this not learning, but a distillation “attack”. It’s an attack on the business model.

Probably the code of these games leaked in some dark corner of the web in the past and that’s why AI can generate it so nicely. AI is also extremely good at building Quake 3, including Carmack’s notes.

But as soon as you are just a tiny bit off the beaten path, Fable/Sol will go postal regularly. So if your indie game has interesting gameplay mechanics, nobody will be able to 1-shot that for a long time. (Because their prompt would have to include everything that’s unique to your game.)

This is fundamentally about price. Does every service use EC2? No! Because at certain scales, the absolute price differences become too large. And then you celebrate the millions in additional profit generated by moving from EC2 to bare metal. It’ll be the same with SaaS: those companies keeping things AI-free will have higher profit margins. So the market will reward AI-free.

Location: Kiel, Germany (CET)

Remote: Yes. EU hours; US-East mornings work. On-site in northern Germany.

Willing to relocate: No

Technologies: C++, Python, PyTorch/TensorFlow, CUDA/SIMD; computer vision (retrieval at scale, inspection, SfM/SLAM); German speech recognition on-prem; real-time audio and kernel drivers; embedded/edge inference; RL environments (PPO, verifiable rewards); the boring enterprise plumbing around all of it

Résumé/CV: https://hajo.me/cv/

Email: [email protected]

I'm the fractional ImageRights CTO, looking for a 2nd startup that needs hands-on tech, scale, or AI help.

I have 18 years of experience across 5 companies in taking products from idea to paying customers. I've built ImageRights' web search and image-retrieval database (C++, in service since 2015, finds copyright infringements worldwide), DeutscheKI (100% private German AI and local ASR), and before that the Spatial Audio Designer, used in Hollywood. And a German image distribution start-up.

I have experience with AI research, adaptation, and training. I've built AI RLVR training tasks in the past. Won the 2021 Bomberland mid-season with PPO. I'm also open to northern German on-prem CV/ASR work in case you cannot ship data to any cloud.

Side interests: I built a Pump It Up pad with custom self-developed and self-manufactured sensors. Currently doing CAD design and policy training (with PPO) for a robot.

Recently I've been debugging MSFS2024 inside Proton: https://github.com/ValveSoftware/Proton/issues/8255#issuecom...

if he “had been made aware that OpenAI has scraped and trained on information that was behind a paywall,” the company would have required OpenAI “to retrain its models.”

... yeah, sure.

I got 3x Bose QuietComfort 35 2 by buying 1. They did honor their warranty, but my impression was that they plan for people not going through the effort of actually requesting a replacement item, even though the warranty says they can.

The first one, "Henckels", openly says that their stuff is from China.

But the second link, "Zwilling" says "Approved". But boy was I disappointed that their "Zwilling Peak" branded ceramic hybrid frying pan for $100 says "Made in China".

"Why is it such a big deal?"

It's used by OpenAI.

I would like to run a robotics company. But raising for that is super challenging, because you have high execution risk combined with low margins on the final product. It's like the opposite of what western VCs want to see. That's why this type of company is usually founded in China.

And I disagree with you. I'm pretty sure parents worldwide would love a clean-up robot. You can avoid a lot of the risks by making it work only for empty rooms, which is OK because you need to clean up and let the vacuum robot run anyway (when the kids are away sleeping in their bedroom). There's been multiple attempts at building DYI robots to vacuum LEGO bricks off the floor. (Because stepping on them at night is such an immediate pain point.) Unless your product is soaked in hostile behavior and surveillance tech, parents will probably be happy to buy an in-home cleaning robot, no PR needed.

I think the main issue here is that parents are typically busy, which makes them bad startup founders, which is why they cannot "scratch their itch" and turn it into a business.

=> home cleaning robots do not exist because western VCs don't like it and parents aren't usually startup founders.

Also, didn't Roomba already leak non-consensual bathroom videos?

"VCs in the US are too big now — only social media (prev) or Al (now) have the scale they want."

I agree.

(I don't work at Anthropic, but I've designed RLVR tasks)

My impression is that especially for long-horizon tasks like science, the harness is much more important than people give it credit for. Claude Code + Fable 5 seems to have a tendency to "give up", get stuck in a dead end, or claim things to be impossible. But using the Fable 5 API together with a custom harness, it'll happily try 200+ variants and fail its way towards the goal.

If you give the AI a way to give up, eventually it will. If you remove that option from the harness, then thanks to the non-determinism inherent to LLMs, you get to explore pretty much all related solution attempts.

That’s what big trucks use. Because you can use wide angle cameras and actually see more.

aggressively proactive

I'd say that's the correct way to describe frontier models. They were trained with reinforcement learning based on human feedback. And obviously, humans prefer the bug-free variant. That's why models are now super verbose and spam tests like crazy. In their training environment, tokens were effectively free. And the humans that got asked never saw the price. If you ask people to choose the better offer and both are free, you end up with bloat. It's like people over-filling their plate at a buffet, then leaving leftovers. Except in this case, it's AI models burning through your wallet.

In my opinion, the big issue with Fable is that Claude Code cannot use it properly. I know, that sounds weird, but I've had Fable run down the wrong lane (and never stop) or give up and claim that something was impossible so many times (until I pointed at a GitHub repo that solves the "impossible" issue).

But a while ago, I had access to a Fable harness that just never gives up. And that verifies itself. It burned $100 in API tokens in 15 minutes ... but it succeeded for all the prompts where Fable + Claude had failed.

And I believe that's a real issue for Anthropic. Fable+Claude is not too expensive thanks to the subscription, but Claude severely nerfs Fable. To save money, I guess. Fable API + Custom Harness is a different class, it's so much better. But API tokens are so expensive, you're cheaper off hiring a freelancer.

My experience with Claude is that it suffers badly from “not invented here” syndrome. So probably it rebuilt something like llama from scratch and then 2 hours suddenly seems reasonable (if you don’t question the approach). And that’s the thing, someone with no experience isn’t going to question it.

"using Codex for inference" says the article

<s> I'm glad they are completely not creepy for a company that plans to make money by selling your private data. </s>

"Capture the daily moments that define you. Voice the hidden details and tell the story only you can tell."

"We may sell, transfer or otherwise share [..] including personal information"

Nuclear Toothpaste

https://en.wikipedia.org/wiki/Doramad_Radioactive_Toothpaste

“many of Hollywood’s problems can be traced back to a single economic theory of everything: the rent is too damn high.”