wuschel

I had the pleasure to speak to your former COO - so glad your organisation exists and publishes its amazing work!

I did not try it, and I am nervous the day I need to - but I think there are some solutions to export photos from the Apple technology stack using open source software [1,2,3].

I remember seeing a longer blogpost re the topic of retroengineering the Apple phot sync story, but I could not find it.

[1] github.com/rcarmo/PhotosExport [2] https://github.com/craigtrim/icloud-photo-export [3] https://icloud-photos-downloader.github.io/icloud_photos_dow...

The main difference total other laptops of non-Apple make is/was unified memory (graphic VRAM + RAM) architecture. No need for an extra dedicated graphics card to get 64+ GB VRAM.

> But IKEA is more often than not more than sufficient quality.

I am not so sure about that. There are some product lines that are somewhat OK price/quality e.g. IVAR seems to be OK. But most IKEA products are too much optimized optimized on the cost side for my taste. The good days of "wooden furniture" is over.

Ha, cynicism indeed. I know where you are coming from.

There is a base level of quality a certain kind of customer is expecting in a product before considering it to be worthless trash.

Let's stick with the cynical position and assume that the code has to run, at least do what it is supposed to do, some time on Monday, part of the time.

When reading posts like this, I am really curious how organisations control for quality.

With an 10x amount of output you 10x the amount of failures at constant production quality. Did you put improved processes into place?

Makes sense re Deepmind. I also guess it is easier to train an AI using Deepmind approach on the Starcraft 2, Go, or other more closed systems than in the rather open business world with all those rules that need to be interpreted by a human judge in case of problems.

Good point re expert systems. I need to look up what actually made them fail, or how they are not scaling.

Have you been working in the space?

> My favorite part of AMT is always going to be figuring out that if we only paid in $0.07 intervals,

That is hilarious! Was it a semi-random discovery due to interaction with the system and people, or did you intentionally looked to game the algorithm?

I am actually with you here. Such a system probably needs to be battle tested in a specific domain of business to see where the deficiencies at this stage are.

The way you have expressed it there are optimized workflows not only the managerial/executive, administrative, content producing set of workers, but in almost every profession, encoded in beaurocracy, checklist, best practices and culture: doctors, care workers, lawyers, policing ...

The question is how well language models have internalized the optimal decision pathway and mitigating strategies to deviating circumstances. What data do they have to be trained on? If an AI defeats the best Go players in the world, surely it is a good question whether they can exceed in games with large unknowns such as business administration.

As mentioned somewhere else in this thread, the moral and ethical layers is where some of the challenges of these system can be found.

I have no experience with data on bad and good CEOs depending on their working context.

At this stage, this seems to be a demo, albeit an interesting one. I find the concept still very intriguing, as it is exactly the thing that attacks white collar functions in administrative and executive tasks sets. I need to give it a spin.

I know. It hurts. But I have the feeling that you get what you pay for.

The purchase cost of H100 or B200 systems with comparable VRAM is a one order of magnitude higher. Although I can only guess how much lower the token/sec output of the Mac Studio will be. Probably 2-3 magnitudes lower?

While a cluster has to work with many users simultanously, and is a good investment for a company, perhaps the Mac Studio will be a good use case for a personal larger LLM deployment configuration.

Perhaps someone has the token/sec numbers for larger models running on older Mac Studios?

> the most productive people with AI on my team are the two youngest. They don’t have product wisdom, yes, but the gap is so noticeable

Is it Software or another problem domain? What problem are they attacking, what stack/method are they using?

> AI isn't adding much productivity

Would love to know where this comes from. I am not sure about that. Could you point to data/anecdata?

I had the same thought.

I hope the other providers will add a geographical limitation on this EU rule.

(On a side note, I wish they would replace those EU beauracts with LLms).

Where is the problem with using LLM generated text?

You could use your own hypothetical house elf to do it for you, or pay someone to do it. LLMs are just cheaper for a certain set of problems.

People will find ways to circumvent this, so this limitation will only hit the technically less adept people.

I was not trying to say that statistical can't generate deterministic code - I was wondering what tools I could use to make sure that they catch all the invariants when given a specification of inputs, logic layouts, and outputs. Certainly, feedback loops have to be involved - e.g. look at the lifetimes guarantees the Rust compiler gives - to make sure the answer is refined in iterative steps.

Hey, thank you so much for jumping in and for your elaboration.

> SOC and piece of test equipment might have it's own control language

OK, so apart from code generation the LLMs are providing language integration and data integration.

But how would you control quality, or make sure the LLms does not create some unwordly control code? I am trying to understand if they are using type checking or any other methodology that keeps the LLM generated code in check.

I am not from the hardware field, but I do understand that we are trying to build deterministic mashines. It is quite mindboggling that we are using statistical models for building the test suite, instead of, say, permutating through all critical input states and testing the output parameters for it.

I am asking because this isnt certainly the only mode of operation for LLMs in testing and quality control, and I would love to get a better understanding how to build such a system.

I understand that they used Claude to create some sort of verification code and test environments from circuit-design information, verification programs, and protocol specifications. But isn't precisely this quite risky due to the statitical nature of LLMs? I would have thought that they rather operate with very rigid test suites based on such software like F*, Coq, and whatnot.

Can someone with understanding of the process chime in here?

Here is the text snippet from the original korean article. Indeed the information was misrepresented:

"Customer-specific SoC verification work, which Usually takes Over a month, was completed in just two days. A second-year employee completed a USB model development that previously takes a month in a single day".

How come Deutsche Bank, a German bank, is the first bank doing this? And why now? Is it because the less friendly political stance towards Europe from the US side that the tables have turned a bit?

I also wonder how much political backing a bank needs to offer a service with the implications towards the status of an allied currency as reserve currency status - small as it may be for now.

> Now 80 percent of it is done using the alternative.

Did not follow the industry since 2018, but it is quite a feat switching the supply chains to the recombinant version. Good to hear that they managed to do it. I wonder how other Pharma and Medtech players are doing in that regard.

The horseshoe crab story comes up on HN now and then. Typical media game, the crabs crawled out their appointed time slot in the media publishing calendar.

Here is a full disclaimer (from the trenches of startup land): I co-founded and ran a biotech startup that tried to disintermediate that particular supply chain. We even flew to SF to speak to some of the YC chaps, and talked to a strategic buyer. Our technology R&D failed to deliver, unfortunately. And the market is fairly small. But I learned a great deal re that particular business, and a lot of more general lessons. All in all, a crazy run with so many crazy stories, I need to write them down one day.

Horseshoe crabs are truly fascinating animals in respect to their physique, lineage and behaviour. But these things one can read up on Wikipedia.

We had quite a bunch of them in our lab (and living room). They were supposed to be voracious predators to all things they can eat, but ours developed a rather expensive habit in captivity by only eating the most expensive of goods from the local fish market. They would not eat anything else.

A pin? I did and fought for the cause. Happy to share. See my other post in this thread.

Sounds interesting, indeed. Just dug up this review [1] that claims there were no adverse effects from any of the interventions, and that after 2.5 to 3 years of use, a fluoride toothpaste containing 10% xylitol may reduce caries by 13% when compared to a fluoride-only toothpaste.

I wonder how well bacteria will adapt long term in your mouth to Xylitol exposure through bubble gum based adminstration.

[1] Xylitol-containing products for preventing dental caries in children and adults. Cochrane Database of Systematic Reviews 2015, Issue 3. Art. No.: CD010743. DOI: 10.1002/14651858.CD010743.pub2.

"Die Gedanken sind frei..." - can someone guess them? A truly dystopian question.

It is been 10 years ago when I looked through some research in that area. Back then it was the problem to get enough good signal. Has the situation changed radically here for non-invasive electrodes or portable detectors?

I understand that the approach would be to turn weak signal, context, and LLM based signal processing into useful human computer interaction. For everything else, I have the feeling we are still in speculative territory.

I love the cool tech demo! I also would have loved to see a offline sensor calibration package - or whatnot - and analysis of model precision instead of Klingon. Just trying to think up of an actual use case where you would actually truly need an LLM instead of one of the other well known data analytics methods.

Aw, that is truly annoying...

Thanks for pointing it out!

Until now I never experienced something like this.

I understand that you are alluring that maximum training-state memory, not parameter count, is the key variable here. So you better start with the smallest model, and go to for highest training data quality, with the outlook of coupling systems together?

Books that come into my mind that might be useful to you:

A Philosophy of Software Design — John Ousterhout

Designing Data-Intensive Applications — Martin Kleppmann

Domain-Driven Design — Eric Evans

There might be some courses as well, but I never looked into them.

This also works quite well:

Write down the problem. Think real hard. Write down the solution. (Thank you, Feynman).

Hi!

Many thanks for the link! I want to make a side project involving some sort of simulation, and I certainly have a look at it for inspiration.

Cheers!