achille

no profile record

Submissions

Thermodynamic Computing from Zero to One

17 pointsby achille8 miesięcy temu3 comments

comments

11 dni temudiscuss

This is amazing! Most definitely headed in the right direction. The separation layer between front and back must be cut at the smallest possible 'slice'.

Lots of people here snarking would understand if they 'felt' the latency and additional overhead. Not enough thought has been put in carfully slicing the data for individual use cases.

I'd go even further, in his demo of 'generating load by moving the config often' -- I think that 'top' app should have 'jit-ed' more of the rendering on the client such that the only information traversing pi<>client is compresed delta's of the ps hose.

2 miesiące temudiscuss

latency on the gBus was at the time ~250msish but would regularly spike much more on certain sections. Each keystroke would have to roundtrip before appearing on the screen.

Using something like mosh would mitigate it, but there was no UDP tunnel available through beyoundcorp.

Cider and MDProxy were running rendering locally so the keystrokes were nearly instant.

2 miesiące temudiscuss

Yep, I made my own! (Xoogler 2017-2023) this is my noogler IDE story, one of my favorite, proudest hacks!

I developed a fork of the IntelliJ IDE on my second week at google out of raw frustration over latency. At the time I was commuting 2-3hrs/day SF<>MTV on the gBus.

Connectivity on the bus wasn't optimal, and there was high latency. Cider didn't have deep integration, and wasn't able to let me explore and understand the internal APIs effectively. I found it easier to enter a debug session within Intellij then 'vibe' and explore the internal apis via superComplicatedObject.ini<tab>.

Faced with an alien architecture + ADHD-unfriendly flow-crushing remote desktop latency -- and the lack of discoverability, I started hacking at it and without any knowledge of the system and architecture. Just tracing Intellij execution, subprocesses and network calls.

I was able to hack together a prototype in a few days that allowed me to run IntelliJ on my Mac, while the heavy bits ran on my corp desktop. The system would mount the remote filesystem over sshfs, would monitor and patch network connections and setup transparent shim binaries. Half of Intellij was running on the Mac (the front end) and the other half ran on Linux. Intellij didn't "know" that that it was running on a mac. This was initially implemented in a ~250 line shell script that patched everything.

It was called MDProxy[1] and ended getting adopted and supported during COVID as more development went remote. This became a source of many peer bonuses and spot bonuses. circa 2017* remote coding options at the time:

         typing   | code
         latency  | integration
         --------------------------
 cider   low      | meh 
 mdproxy low      | great
 ssh+vi  med      | meh
 rdp+iJ  crushing | great

[1] https://github.com/bazelbuild/intellij/blob/6b8f03c21172033a...

3 miesiące temudiscuss

What will happen to my "Grandfathered Plan" I signed up to test it, don't recall if I gave you my credit card

3 miesiące temudiscuss

url is 404'ing; what is the source for the skateboarding clip?

3 miesiące temudiscuss

what setup do you use for the bar at the bottom?

4 miesiące temudiscuss

due to library limitation, not language, they could use fancy-regex

4 miesiące temudiscuss

i registered for their demo and it seems to be unable to do any of the advertised features. nearly everything crashes; feels like a ux wireframe not yet wired it

4 miesiące temudiscuss

same here, would love to compare notes

4 miesiące temudiscuss

  - local data residency & sovereignty
  - latency
  - bandwidth 
  - regulatory climate
  - competition

uae is business friendly

all cloud providers have middle east presence

refineries generate terabytes of sensor data per hour

the population and people there produce and consume a lot of data

4 miesiące temudiscuss

Fun challenge: I asked Claude/Gemini to decode the audio by just uploading it as puzzle.wav. Claude is able to decode it:

https://claude.ai/share/4262fb6b-3ca1-407f-af0d-4d014686e65d

5 miesięcy temudiscuss

in the article they explicitly said they stripped symbols. If you look at the actual backdoors many are already minimal and quite obfuscated,

see:

- https://github.com/QuesmaOrg/BinaryAudit/blob/main/tasks/dns...

- https://github.com/QuesmaOrg/BinaryAudit/blob/main/tasks/dro...

5 miesięcy temudiscuss

Absolutely, they didn't give the agents autonomy to research or any additional data. No documentation, no web search, no reference materials.

What's the point of building skills like this?

6 miesięcy temudiscuss

Spoilers: https://areweseductionyet.pages.dev

6 miesięcy temudiscuss

Nice, that worked. It's even faster.

                 yours  yours+opt  claude
  ---------------------------------------
  Time (ns)        40.9      36.4    38.7
  sqrt/iter           3         2       1
  Instructions      207       187     241

Edit: it looks like the claude algorithm fails at high eccentricities. Gave chatgpt pro more context and it worked for 30min and only made marginal improvement on yours, by doing 2 steps then taking a third local step.

https://gist.github.com/achille/23680e9100db87565a8e67038797...

6 miesięcy temudiscuss

thanks for sharing that, it was simple, neat, elegant.

this sent me down a rabbit hole -- I asked a few models to solve that same problem, then followed up with a request to optimize it so it runs more efficiently.

chatgpt & gemini's solutions were buggy, but claude solved it, and actually found a solution that is even more efficient. It only needs to compute sqrt once per iteration. It's more complex however.

                   yours  claude
  ------------------------------
  Time (ns/call)    40.5   38.3
  sqrt per iter        3      1
  Accuracy        4.8e-7 4.8e-7

Claude's trick: instead of calling sin/cos each iteration, it rotates the existing (cos,sin) pair by the small Newton step and renormalizes:

  // Rotate (c,s) by angle dt, then renormalize to unit circle
  float nc = c + dt*s, ns = s - dt*c;
  float len = sqrt(nc*nc + ns*ns);
  c = nc/len; s = ns/len;

See: https://gist.github.com/achille/d1eadf82aa54056b9ded7706e8f5...

p.s: it seems like Gemini has disabled the ability to share chats can anyone else confirm this?

6 miesięcy temudiscuss

The analysis and tools were most definitely ai-aided, but this was done with homegrown forensics tooling, and about an hour of labor that involved cc, gemini (to check for synthid), chatgpt + a lot more. I also signed up for ideogram and generated more images to try and replicate the output: - e.g: https://ideogram.ai/g/fcqp-qTlQV-moS-OuhebhQ/1 (although I refused to pay for ideogram so I could not get a png output, only jpg)

6 miesięcy temudiscuss

I ran some analysis on the source image — this is almost certainly AI-generated. In addition to the visual markers others noted (no photography in federal court, nonsensical positioning), here's what comes up in the original image:

  Filename: ..._simple_compose_01kdcxamjmekery2m9tay43szn.png
    - "simple_compose" + LSB common (e.g ideogram) output 

  Resolution: 1536x1024
    - Exact native output of GPT-image-1, Gemini/Imagen, Flux models

  PNG encoder fingerprint: 0x78 0xDA | single IDAT | 94.7% Average filter
    - Matches PIL/Pillow with optimize=True

  Steganographic watermark:
    - LSB entropy: 3.0/3.0 (maximum)
    - Bits 0-3 of RGB channels filled with encrypted payload
    - ~1.77 MB of pseudorandom data embedded

7 miesięcy temudiscuss

Could you share your skill and workflow? does claude launch codex in a tmux session?

7 miesięcy temudiscuss

Paper has some more useful examples:

https://imaging.cs.cmu.edu/svaf/static/pdfs/Spatially_Varyin...