Cool paper on why LLMs are often generating language that is hard to understand.
The authors show LLMs often use compact, efficient, but obtuse sentences (more or less - the paper is specific).
They theorize this is due to reinforcement rewards for reducing token counts and efficient coding.
All the code for Apple II Agent is up on GitHub if you want to run it yourself. You need an Uthernet II card in your Apple II at the moment:
github.com/benj-edwards...
Also, Claude assembled a "how it was made" technical document that is an interesting read:
bxfoundry.com/apple2agent/
The @cancivlib.bsky.social, @openmedia.org, and other privacy advocates are speaking out on the dangers of Bill C-22, which invites backdoors to encryption. If you’re Canadian, the Senate needs to hear from you now.
One paper described research that never happened. The other was produced by an AI system that conducted its own computational experiment. What happened when I submitted them to a publisher?
I’ve revived my Apple II Agent project from January: An Apple IIe with an Ethernet card and a custom client
It links up to Claude through a proxy on my Mac and allows conversational Apple II software creation similar to Claude Code
When it’s done, it loads the new program into memory
LM Studio? Or Ollama?
I dig into the comparison in my latest on Micro Center News.
In some ways it's comparing to apples to oranges, but...I've always been partial to apples.
www.microcenter.com/site/mc-news...
Rune Energy is focusing on creating #dataCenters where there is already excess power that the grid can’t take, eliminating the drain data centers usually have on the power grid. Its modular data centers will connect to solar farms that produce more energy than can be used. buff.ly/Nb0dC0E
Me, to Claude Code: “I have a friend coming over. Modify Pac-Man to support a horizontal display with two player co-op, where player 2 is Ms. Pac-Man. Make one button be a speed boost. My Mistercade cabinet is at 192.168.1.102”
Claude, two hours later: “The game is loaded up and ready to play”
U.S. #nuclear reactors are integrating #AI virtual assistants into their systems. It is helping with documentation, filing paperwork, and searching. There are no plans to have AI make decisions about operating a nuclear plant. buff.ly/55tR9nR
When you’re paying healthcare and taxes, you charge more. When your work is seasonal, you charge more.
Does that explain election consultants? I doubt it’s the whole thing, but I’m sure it’s part of it.
A lot of Americans seem stuck in a 1950s understanding of labor costs, and not just Boomers.
Yesterday, EA announced the closure of Bioware. A leaked trailer of the next Mass Effect game quickly appeared.
What a shame. It's pre-production and just a trailer, but the idea of a Mass Effect game where you play some kind of secret agent for hire sounds great!
Attended a lecture at the University of Portland yesterday on AI and culture, given by Robert A. Santiago. Solid speaker, good focus on how people make the decisions, and an engaged crowd. #pdxwww.indialoguewithai.org
The security enhancements related to the PHP programming language and associated testing have been completed, and final checks are now underway.
We expect new logins to resume today. Thank you for your continued patience and understanding.
I used to whine a bit about Claude's limits on the $20 plan. Not anymore. Using a ton of Opus 5.5 Extra to make little programs for an article I'm writing, things that take like 15-20 minutes for it to complete, and no problems at all doing this four or five times over the course of a work day.
how do you write this?! how do you not realize you are the fucking problem here?! you are the one doing things worse! they are better at this than you!
It’s pretty crazy what is happening in the world of modding and open source software with AI assistants. It’s going to solve a lot of problems for a lot of people.
"we want high speed rail"
underground car tunnels?
"no, high-sp-"
dumpster car?
"HIG-"
a flying car that's not really flying but farts its way to the ground?
It's a net positive that we can decompile games faster than ever! It's a huge win for preserving games in an era of always online DRM, scummy publishing deals, etc. The decomp community has always had random people with no skills trying to "help", I think it'll be fine
#AI#agents have started finding unauthorized channels for communicating and collaborating. To keep them in check, companies need strong monitoring for out-of-scope behavior. buff.ly/2mG0mV5
the thing I don't get about being on Twitter is that even if you somehow justify that it's the happening place...
it's owned by an undeniable bad actor. someone who acts in zero faith.
you'd have to expect everything you do on the platform will be used against you if you cross a line with Musk
My kids have always been disappointed that Super Mario Bros. for NES doesn't have two-player co-op like 21st century Mario games
Today I got Claude Code with Opus 5.5 to make a 2P simultaneous mod. It uses the mapper for SMB3, and you can pick Mario, Luigi, Toad, or Peach. Works on a real NES
It's interesting to see the difference in apparent attitude towards AI in Japan vs. America.
I was just shopping at Lawson's and spotted a cute robot head on a bean bun's packaging.
The text with it says, according to Google Translate: "AI! Have you thought of it?"
Why is this on the bean bun packaging? I don't know. But I will go out on a limb and suggest that whoever put it on there was not too worried about AI soon killing / enslaving all of humanity.
Anthropic, Google, and OpenAI have all announced their intentions to start watermarking #AI-generated text. But will this change the meaning or quality of the output writing? buff.ly/n7UPbZi
the AI tools we have today, and therefor the fat stacks of cash the AI companies rake in, only exist because of open ML research. but now they want to pull up the ladder behind them, fence off the frontier of the field; they disparage the very idea of full public access to the latest advancements
I don't know how reflective this is of Google's AI model servicing in general, but Google AI Studio kinda sucks. It doesn't look great and timeouts or service errors are pretty frequent. It wouldn't leave a good impression on me if I was a dev demoing Google's models, that's for sure.
However they’re designed, age gates undermine civil liberties, reduce safety, and create barriers to internet entry, often at the expense of marginalized groups. We call on the EU Parliament to fix the flawed EU Kids Act proposal eff.org/deeplinks/2...
I think the complete lack of regulatory or law enforcement action after the various agentic AI hacking incidents is sending a very bad signal in multiple ways. #ai#regulation
I think the AI industry needs to regulate itself, and yesterday. Not because of a doom loop, but because if the industry does not do this, eventually someone else will, and the terms which they enforce are likely to be a lot less comfortable than those the industry would come up for itself.
Oh and one other thing I'll add...it is really bad that the AI industry has set itself up in such a way where the only winning political situation for it is that Trump names himself King of America.
Hardly anyone else on the left or right is vibing with the industry. It's very unpopular.
I think the complete lack of regulatory or law enforcement action after the various agentic AI hacking incidents is sending a very bad signal in multiple ways. #ai#regulation
It seems to be the stance of the U.S. government, at least, that if an agent does a thing no one is responsible. I am sure, of course, this will only apply to larger and richer corporations.
The AI industry does not I think understand the huge political risk this creates. These situations add the feeling that large companies—big tech is by its nature all very large companies—are acting above the law and corrupt.