physix

I'm curious to know the business model. I assume you charge the companies, but I couldn't find the pricing model on the website.

Are you able say something about this?

BTW: The AI Assistant is itself a Cyoda app.

Do you have any examples?

I was waiting for

          .....   ......   ......   ....   ..    ..   ......
         ..       ..   ..  ..   ..   ..    ...   ..   ..     
         ..       ..   ..  ...  ..   ..    .. .. ..   ..     
          .....   ......   ......    ..    .. .. ..   ..  ...
           .....  ...      .. ..     ..    ..   ...   ..   ..
              ..  ..       ..  ..    ..    ..    ..   ...  ..
         ......   ..       ..   ..  ....   ..    ..    .....

to show up, until I realized I've been coding too much today.

me too

I was thinking more of the former, whereby I "bring my own servers".

I haven't really thought it through yet, whether that even makes sense.

This is a great idea. I really like it!

We considered reaching out in May, but held back because we want to run on bare metal.

Any chance to get this provisioned on bare metal at Hetzner?

We have K8S running on bare metal there. It's a slog to get it all working, but for our use case, having a dedicated 10G LAN between nodes (and a bare metal Cassandra cluster in the same rack) makes a big difference in performance.

Also, from a cost perspective. We run AX41-NVMe dedicated servers that cost us about EUR 64 per server with a 10G LAN, all in the same rack. Getting the same horsepower using Cloud instances I guess would be a CCX43, which costs almost double.

Ah, yes, of course! Thanks.

Even when the distance between the centres of mass of two colliding galaxies become comparable to their size?

It's a long time ago, but what I remember was being fascinated by the shapes of the galaxies emerging from a collision under this centre-of-mass approximation, and that it created shapes we see out there. It was as if the main effect were a central mass in each galaxy dominating the dynamics.

This reminds me of when I was a physics undergrad way back in the mid 80s. We used to spend nights drinking beer and hacking some simulations from the Computer Recreations section of Scientific American.

Once we wanted to simulate the dynamics of galaxies. I don'it think it was an SA article, but we did it the slow way by calculating the force on every star individually from each other star. It was excruciatingly slow and boring.

Then some time later, I don't recall where I picked that up, I updated the simulation to just model the force on each star coming from the galaxy's centre of mass.

I could simulate many more stars, have galaxies collide and see them spin off with their stars scattering around.

What struck me was that they looked like real galaxies we see out there.

I wasn't aware of the postulations made in the 60s/70s about there being supermassive black holes at the centre of galaxies, but to me, this simplified simulation was kind of like a smoking gun for that... from an 80286 IBM PC AT.

What's so evil about Netflix?

They use Cassandra and make cool series ever now and then, like Love Death Robots. :-)

It's probably too much work, but it would be nice to see a short comment on the "curated" examples to better understand the reasoning behind the assessment. Why was it included ? What was particularly good about it? That might help people choose the right ones for their use case.

When I lived in Germany, in the 90s, I regularly sat in diesel Mercedes Benz taxis with over a million kms under the hood. Private drivers usually. Many had giant mileages.

We used to say (tongue in cheek) that after 250k, the MB diesel engine was broken in. I don't think MB makes them like they used to anymore.

About 1: Indeed. The moderator remarked at the end that once the interview was over, Dario's expression sort of sagged and it felt like you could see the weight on his shoulders. But you never know if that's part of the act.

About 2: Ah, yes. So if one vendor gains sufficient momentum, their advantage may accelerate, which will be very hard to catch up with.

I think this is actually good, because it means there is no clear winner who can sit back and demand rent. Instead they all work as hard as they can to stay competitive, hopefully thereby accelerating AI software engineering capabilities, with the investors footing the bill.

In the way back days when we submitted our CVs on paper, I always cut mine to a smaller size than letter, in a branded folder. People tend to stack things with the smaller items on top. I don't know if mine actually was on top of the stack, but I can say that I basically always got the contract.

For those who happen to have a subscription to The Economist, there is a very interesting Money Talks podcast where they interview Anthropic's boss Dario Amodei[1].

There were two interesting takeaways about AGI:

1. Dario makes the remark that the term AGI/ASI is very misleading and dangerous. These terms are ill defined and it's more useful to understand that the capabilities are simply growing exponentially at the moment. If you extrapolate that, he thinks it may just "eat the majority of the economy". I don't know if this is self-serving hype, and it's not clear where we will end up with all this, but it will be disruptive, no matter what.

2. The Economist moderators however note towards the end that this industry may well tend toward commoditization. At the moment these companies produce models that people want but others can't make. But as the chip making starts to hits its limits and the information space becomes completely harvested, capability-growth might taper off, and others will catch up. The quasi-monopoly profit potentials melting away.

Putting that together, I think that although the cognitive capabilities will most likely continue to accelerate, albeit not necessarily along the lines of AGI, the economics of all this will probably not lead to a winner takes all.

[1] https://www.economist.com/podcasts/2025/07/31/artificial-int...

Started reading it. It's really fascinating topic. But as an outsider it's quite difficult to assess its potential.

How far away are we from seeing a real application of this in the AI space?

(It reminded me of an ex-colleague's company in Germany that makes analog computers https://anabrid.com.)

Its overuse of the word "comprehensive".

Probably will get worse over time as it ingests all its AI generated material for the next version. Soon everything will be comprehensive.

Can someone put this into context? If I sign up via github, I will be allowing the tool to "act on my behalf", with write access to my profile.

Why is it asking for write access to my profile?

I had Google Gemini 2.5 Flash analyse a log file and it quoted content that simply didn't exist.

It appears to me like a form of decoherence and very hard to predict when things break down.

People tend to know when they are guessing. LLMs don't.

A game changer in which respect?

Anyway, this makes me wonder if LLMs can be appropriately prompted to indicate whether the information given is speculative, inferred or factual. Whether they have the means to gauge the validity/reliability of their response and filter their response accordingly.

I've seen prompts that instruct the LLM to make this transparent via annotations to their response, and of course they comply, but I strongly suspect that's just another form of hallucination.

I like the planning bit, though.

Talking about hilarious, we had a Close Encounter of the Hallucinating Kind today. We were having mysterious simultaneous gRPC socket-closed exceptions on the client and server side running in Kubernetes talking to each other through an nginx ingress.

We captured debug logs, described the detailed issue to Gemini 2.5 Flash giving it the nginx logs for the one second before and after an example incident, about 10k log entries.

It came back with a clear verdict, saying

"The smoking gun is here: 2025/07/24 21:39:51 [debug] 32#32: *5902095 rport:443 upstream timed out (110: Connection timed out) while reading response header from upstream, client: 10.233.100.128, server: grpc-ai-test.not-relevant.org, request: POST /org.not-relevant.cloud.api.grpc.CloudEventsService/startStreaming HTTP/2.0, upstream: grpc://10.233.75.54:50051, host: grpc-ai-test.not-relevant.org"

and gave me a detailed action plan.

I was thinking this is cool, don't need to use my head on this, until I realized that the log entry simply did not exist. It was entirely made up.

(And yes I admit, I should know better than to do lousy prompting on a cheap foundation model)

Read the comments here so far and I find that they are absolutely right to offer an AI layer that speeds up building apps on their db.

Once built, the solution is plain-old-runnable-code (PORC :-), as long as the business logic implemented doesn't exit to LLM. So I don't fret so much about the AI hype story here.

For anyone starting off building with new tech, an AI assistant is really helpful.

What are they going to spend those $2B on?

My favorite at the moment is the Augment Code plugin for IDEA and PyCharm, because it gives me time to read HN while waiting!

nope

This looks to me like the smoking gun on a type of acquisition that circumvents regulatory oversight, primarily driven by the "need for speed":

https://medium.com/@villispeaks/the-blitzhire-acquisition-e3...

which I first saw here

https://news.ycombinator.com/item?id=44553257

Actually I didn't. Correct me if I am wrong, but my understanding is that RL is still an LLM tuning approach, i.e. an optimization of its parameter set, no matter if it's done at scale or via HF.