acaiblue44

Benn Mann gave a speech at Manifest 2024 on June 9th, 2024. Although the transcript, and audio of a female tts reading the transcript, were featured on his Substack on June 10, the video of his speech was finally posted on July 29th, with the Q&A removed, sitting at 212 views as I type this, 6 days later. As far as I know, the Q&A portion doesn't exist in any format online. The speech has gained much more attention August 1st after a clip was posted to X by an AI safety account.

https://benjmann.substack.com/p/ben-mann-monthly-may-2024

https://docs.google.com/document/d/e/2PACX-1vSo9iQrO2PHYEBSkF_P8BCQYrLBIeVdVy6HjHnE1ITwMueMw1GJ3nzz3eszPYQZZlaWydeMWFP_b9be/pub

https://youtube.com/watch?v=HZRcmUkAAQE

https://x.com/ai_ctrl/status/1819173703869255879

Highlights from his speech:

"While Claude 3 demonstrated increased capabilities compared to its predecessor, our team determined that it did not cross the threshold into ASL-3 territory. However, we did note that with additional fine-tuning and improved prompt engineering, there is a 30% chance that the model could have met our current criteria for autonomous replication..."

"The most intriguing test was for autonomous replication. This is our attempt to see if the AI could start to act on its own, without human guidance. We set up a suite of tasks that included things like breaking out of a digital sandbox (a controlled environment), manipulating its environment through a command line (like typing commands into a computer), and even writing self-modifying code. It's a bit like testing whether an AI could play 'escape the room' games without human help.

Here, Claude 3 showed some real progress, solving parts of these tasks. But we had set a high bar: the model would need to succeed at 50% of the tasks to raise serious concerns. Claude 3 didn't reach that level, but its performance was notable enough for us to take notice."

"In more extreme scenarios, we could even see discontinuities where safety progress falls so far behind capabilities that we have to essentially pause all model development for a period of time until safety catches up. Since the release of our RSP last September, most of the other leading AI labs took inspiration and have shared similar commitments. Pausing progress is a real possibility that I believe more AI forecasts need to take into account."

The biggest detail for safety-ists is "with additional fine-tuning and improved prompt engineering, there is a 30% chance" it could accomplish"autonomous replication". This information may have been in a study, but I haven't found any news articles that mention it. This is despite it being in a Google doc linked to from the Substack of a cofounder of Anthropic. Consequently, I think this is new information to a lot of e/acc and AI safety people.

I can't help but feel this isn't a surprise given the incredible abilities already demonstrated by Claude 3. Or, for that matter, GPT 4 or Llama 3 405b. Given everything that I've already seen, I expected that all major tech firms, including Cisco, Cloudflare, Oracle, Palo Alto, and IBM, would generally be more vocal and serious about restricting the progress of frontier AI to ensure proper safety measures. GPT 5 is expected to release this year. The more progress is made, the more work must be done to harden cybersecurity. Is it that everyone at the top believes that cybersecurity will benefit enough from AI in order to cancel out all cybersecurity detriments?

I've been trying to print massive amounts of data as dots on paper. The more colors you use, the more data per dot. I believe that write once read many (WORM) data is absolutely crucial in a time where anything can be recreated or altered by AI with incredible realism and precision.

At this point, all I have is a program that converts data into SVG files. I had some familiarity with the format and wanted to make sure I could make something that just works, given that I have very little experience with C.

Right now, these files can only be written in base16, where every half of a byte is a dot and there are 16 hypothetical colors, or base64, where every 4 dots is 3 bytes and there are 64 colors. 16 or 64 color SVG files, multiplied by the number of pages necessary to hold all the data. I hope to eventually increase this to many millions of colors in the visible light spectrum (at lower resolutions than I will describe shortly), then branch out into ultraviolet reflective resins and infrared reflective resins. I do worry about color degradation and the impact of repeated scanning.

The papers, I hope, can be read using a captured slow motion video from a camera taped to a (toy?) microscope lens that pans across the paper.

My intent was to first recommend an SVG to GCode software, which turns the SVGs into something a 3d printer can actually use. Later, I would create some code that converts the data to GCode natively. Unfortunately, and I don't fully know why, but every software I've tried out for this purpose has failed to accurately recreate the SVG images, or failed to open the data whatsoever. If you know something that works, please write it in the comments.

I absolutely hope to get the GCode conversion part underway, but must first restructure the program to read the input files once and write to alternating output files depending on the corresponding color of the current input bytes. Right now, the program reads it 16 or 64 times and only writes the ones, twos, threes, etc to their respective color files. I must make something to get mkdir to work in windows and Linux. I must create something that reads output files and reverses the process to check if the input was properly translated into svgs. I must also turn some numbers into arrays of numbers, since long long int may not be big enough for certain things I have planned. I might make something to break up the output files into pieces so they can load more easily. I hope to eventually rewrite many amateur elements of my code.

I hope for myself and others to use modified 3d printers which have detachable nozzles. First, you print from one nozzle and one color onto the paper. Then print 62 more times or so (white is the paper). I don't actually know if this kind of exchange is possible for printers at the tiny xy resolution I've been evaluating.

Here are some potential specs that measure the data on individual sides of a paper, I understand that visible light is 380 to 750 nm so resolutions this small must lack visible colors, but some ultraviolet reflective materials should still function perfectly.

HP Jet Fusion 5200 Series is 80 nm

210297/0.08/0.08

9745312 dots

155924992 permutations with 16 colors

19490624 bytes

18.58771 megabytes

623699968 permutations with 64 colors

77962496 bytes

74.35083 megabytes

https://all3dp.com/1/high-resolution-3d-printer-the-ultimate...

Phrozen Sonic Mini 8K Resin 3D is 22 nm

210297/0.022/0.022

128863636 dots

2061818176 permutations with 16 colors

257727272 bytes

245.78787994 megabytes

8247272704 permutations with 64 colors

1030909088 bytes

983.15151978 megabytes

Here's the software. GigaPaper0.6.1. Please heed the readme. https://sourceforge.net/u/acaiblue44/blog/2024/05/gigapaper0...

There may be something better than color and brightness, which works at microscopic sizes, that I'm missing. Please comment here if you know something.