Monthly Archives: July 2024

Phil 7.31.2024

And so we say goodbye to July. That was fast

11:00 meeting with Wolfram – went well!

SBIRs

  • Back to setting up the NNM project codebase. Here’s what I was planning before getting pulled off:
    • First, a program that reads in a config file and generates a file of activations for each layer. This will be a fixed size ring buffer of tokens so that the number of vectors for each layer is fixed for the “evolution”
    • Then code that reads the same config file to produce a set of UMAP files, again one for each layer.
    • Last, the visualization code using Plotly. It either reads the same config or the user can navigate to it. Management of all the files (12 layers of embeddings, 12 layers of UMAP) will be managed by extension. Might want to set up folders though.
  • Started up the instance and getting back to work

Phil 7.30.2024

SBIRs

  • 9:00 Standup
  • 10:00 AI Ethics talk – good!
  • Maybe some NNM – nope. Couldn’t get settled into a coding groove
  • Moved the the CwoC Overleaf project to the company repo.

GPT Agents

  • More reviews. I have been pummeled by too many words

Phil 7.29.2024

SBIRs

  • Looks like I’m going to Huntsville for a conference
  • 9:00 Intern presentation. Good stuff 🙂
  • Got all the flight, hotel, and car rental handled
  • Looks like book club is on for Thursday

GPT Agents

  • Start journal review – started

Phil 7.25.2024

AC gets its checkup: 7:30 – Noon

SBIRs

  • Finish the whitepaper(s). One is more marketing, one is more proposal. Done. Not done, refining.
  • 2:00 ML discussion? Kind of. Important no shows
  • 2:30 USNA proposal discussion
  • Maybe do more on NNM?
  • Cancelled the book club for the week

GPT Agents

  • 3:00 meeting. I think we’re discussing the invited paper. One of the things I think I’ll suggest is the HST thing that I worked out yesterday. It’s 639 words. And this is also a good thing to look at: From Burnout to Balance: AI-Enhanced Work Models

Phil 7.24.2024

RE tax! – done, and scheduled for Dec 23, too

Put up my first Bluesky thread. Not sure if I did it right

I also wrote another thing! Need to submit this somewhere.

SBIRs

  • Set up the project folder structure and put together an initial config JSON file. Beginning prompt, token buffer size, and the number of steps. Maybe the number of UMAP dimensions? Or do we always produce 2d and 3d versions?
  • To processing tests to CSV files. That way we can look at them in the IDE
  • Got the structure set up and the config file read in
  • Need to do some BD as high priority

GPT Agents

  • Based on the chat with Jimmy, I loaded the (for me anyway) unreadable versions of Das Kapital into Notebook LM and used the following prompt:
    • Extrapolate to concept of workers and machines to generative AI which is trained on the output of writers (the workers) and can be used to produce mediocre writing without workers, and faster, high-quality writing with them. On one side, generative AI can replace workers, but will produce largely unoriginal content that will lose appeal over time. On the other, workers have access to a mechanical co-worker that can accelerate and often improve what they produce. Is the technology of generative AI inherently capitalist? If so, why? Provide alternative, egalitarian perspectives as well. Use the form of a white paper
  • Here’s the response:
    • Is the technology of generative AI inherently capitalist?
      • To address this question, this paper extrapolates from the sources’ concepts of workers and machines to the context of generative AI, which, analogous to a machine, is trained on the output of writers (the workers)
      • Similar to the machines described in the sources, generative AI can produce output (in this case, writing) without the involvement of writers once it is trained on their work12.
      • However, also similar to the sources’ depiction of machines, the output from generative AI is likely to be of higher quality when it is used as a tool by human workers134.
    • Capitalist Perspective:
      • The sources, focusing primarily on the dynamics of capitalist production, offer compelling arguments for why the technology of generative AI, within the framework of a capitalist system, could be considered inherently capitalist:
      • Exploitation of Labor: A core tenet of the sources is that capitalist production inherently seeks to extract surplus value from labor. In the context of generative AI, this dynamic could manifest through:
      • Replacement of Workers: Generative AI could potentially replace writers, particularly in producing low-quality, formulaic content, mirroring how machinery replaced skilled artisans in the sources456. This replacement allows capitalists to reduce labor costs and increase profits.
      • Intensification of Labor: Generative AI, even when used as a tool by writers, can lead to an intensification of work78. This intensification could take the form of pressure to produce more content in less time or to compete with the output of AI, ultimately benefiting the capitalist through increased productivity.
      • Control Over Production: The sources emphasize how machinery, under capitalist control, consolidates the power of capitalists over the production process459. This control extends to generative AI:
      • Ownership of AI: Capitalists who own the AI technology hold significant power. They control the data used to train AI, influence the types of content produced, and determine how the technology is implemented—all of which can be leveraged to maximize profits.
    • Alternative, Egalitarian Perspectives:
      • While the sources primarily focus on capitalist applications of machinery and labor, they also offer insights that can be extrapolated to imagine more egalitarian applications of generative AI:
      • AI as a Tool for Empowerment: Generative AI, when viewed not as a replacement but as a powerful tool, has the potential to empower writers1011. It can:
      • Increase Productivity: AI can handle tedious and repetitive writing tasks, freeing human writers to focus on more creative and engaging work.
      • Improve Quality: AI can offer suggestions, refine language, and identify errors, ultimately contributing to the creation of higher-quality content.
      • Democratization of Writing: One could envision a scenario where access to generative AI is not limited to capitalists. Open-source AI models and collaborative platforms could allow a wider range of individuals to participate in the creation and dissemination of written content, potentially fostering greater creative diversity and reducing the dominance of large corporations in media and publishing.
    • Conclusion:
      • While the capitalist application of technology is a central theme in the sources, it is crucial to remember that technology itself is not inherently capitalist. The sources highlight that the capitalist exploitation of labor arises from social relations, not from the tools themselves1213. Similarly, the impact of generative AI will ultimately depend on how it is developed, implemented, and controlled. If these processes are driven by a focus on collective benefit, shared ownership, and the democratization of knowledge and resources, then generative AI can become a powerful force for a more egalitarian future.

Phil 7.23.2024

I wrote a new thing. It’s kind of based on how a conspiracy theory might be floated, but from a left leaning perspective

SBIRs

  • Overleaf has changed their personal plan that makes link sharing the same as invites. Upgraded my plan, but grudgingly.
  • 11:30 Sprint planning
  • Time to get back to the GPT2 + UMAP – started
  • Had to update my passwords on bastion and driver
  • Remarkably, everything still works!

GPT Agents

  • ICTAI 2024 paper bidding started. Need to do this today, probably – done
  • An interesting back-and-forth with Jimmy.

Phil 7.22.2024

I have done laundry that included a few items from the Luxembourg trip, so I guess I’m officially done with travel for a while!

SBIRs

  • 9:00 Sprint demos
  • 3:00 Sprint planning – delayed?
  • Spent a good deal of time talking to Ron about the interns and setting up the Overleaf project

GPT Agents

  • Review paper for ACM Transactions on Interactive Intelligent Systems

Phil 7.19.2024

Chores

  • Bills – done
  • Lawn – done
  • Clean house – done
  • 11:30 Meeting with Tim – done
  • Groceries – done

SBIRs

  • 9:00 USNA meeting with Ron – done

GPT Agents

  • I’m thinking of writing up something about providing value to consumers vs. extracting value from consumers. How this manifests in modern consumer culture (many examples of both, though timeless fashion vs. fast fashion might be a good one). Importantly, modern AI (in particular LLMs) has an innate (statistical) understanding of each of these value propositions. These systems, like any technology, can be used for either, but in this case, they can help provide a level of reflective insight, both to consumers and producers of technology platforms (private and government). I think it would be possible to generate some proof-of-concept examples along the lines of the browser plugin that I’m starting to work on. Maybe even directly reference romance scams? They are an extreme example of extracting $$  while promising to provide companionship to the lonely.

Phil 7.18.2024

Still muggy, but not as hot:

SBIRs

  • See what’s going on with my expenses – progress!
  • Finish first pass at white paper? I think it’s just copy previous work over from MDBE? – done
  • 9:00 standup – done
  • 9:30 – 11:30 Thunderbolt – done
  • 2:00 CNAS Catastrophic AI webinar

GPT Agents

  • 3:00 – Starting meetings back up (Discuss conferences and upcoming work ATT&CK social and browser plugin)

Phil 7.17.2024

Looks like the last hot, muggy day for a while

SBIRs

  • More white paper. Slow progress, but progress. This will have to slide some because of the thunderbolt thing. Talk to Protima about her part. Maybe tomorrow though.
  • For USNA meeting with Ron on Friday:
  • 10:00 Thunderbolt meeting – done. Added two slides to the deck and a bullet point.

Phil 7.16.2024

Ouch!

SBIRs

  • Write up stories for Social att&ck (For ETF) and WhiteHat AI browser plugin (For IUI)
  • Call about Identrust – done!
  • White paper. Get charge number. One of the things that I think I’ll mention is very low bandwidth loop management. Something on the order of a coordinate in embedding space (probably projected down to a reasonable number of dimensions. This is sent out as regularly as possible so that other model’s predictions of each other can be validated to a degree. If the distance in embedding space is too great, the likelihood that there is a mutual understanding is low(er)
  • 9:00 standup. Get ride in before the heat – done
  • 2:30 AI Ethics – done
  • 3:30 Thunderbolt – done. Two slides by tomorrow COB? Tricky problem. I guess we just have to sound smart? I think cells could be an interesting direction to think about.
  • Write up something for our interns by Friday to go over with Ron. And follow up with the interns on the following Friday.

Phil 7.15.2024

Gradio is the fastest way to demo your machine learning model with a friendly web interface so that anyone can use it, anywhere!

SBIRs

  • Update Jira – done
  • Submit expenses (scan parking!) – done
  • Write up accomplishments – done
  • 1:30 Meeting – done. Not much that we can really do here, I think
  • Ping Jonathan Gratch and Cosmin Munteanu – done
  • Write up stories for Social att&ck (For ETF) and WhiteHat AI browser plugin (For IUI)
  • Restart book club for when Aaron gets back?
  • Call about Identrust
  • Looks like I gotta write a white paper

GPT Agents

  • Ping folks to set up weekly meetings again – done

Phil 7.13.2024

Back from the CUI 2024 conference, which was fun and useful I think the next thing to do is to put together an “exploit table” for human beings that resembles the ATT&CK matrix. That’s one paper. The other effort will be to put together a white-hat AI Chrome plugin to reduce exploits and scams on the elderly. That should allow progress in a smaller domain to see if such an approach can work. I’d say young adults too, but the IRB might be impossible, and the number of platforms is more daunting. I think it would have to go in a phone OS, and even then it might not be able to handle encryption of all the data presented in, say, a TikTok video.

And this is part of that matrix: Russia recruits sympathizers online for sabotage in Europe, officials say

  • Europe has been grappling with an increase in Moscow-led sabotage attacks as Russia turns its focus to increasing the cost of Western support for Ukraine.