Tasks

• Schedule physical

GPT Agents

• Start working out travel to Sidney Done! I accidentally got tix for SDY rather than SYD. One is in Montana. Ooops. Seems like it has happened before

SBIRs

• Get Mors abstract submitted by Feb 10
• Got storing and loading of reduced embeddings and parameters in NarrativeExplorer
• 9:15 standup – done
• 1:00 Biweekly meeting – canceled
• 3:00 New SBIR meeting – meh

Book

• More proofing – done!

Tasks

• 10:00 Eye exam
• Schedule physical

GPT Agents

• Start working out travel to Sidney

SBIRs

• Get Mors abstract submitted by Feb 10
• More GPT Index

Book

• More proofing
• Feeling pretty good about the originality of my cover

OpenAi has been busy. First, they have some tutorials about interfacing with document collections using embeddings. Looks like a simpler version of GPT-Index

Second, they wrote up a report on using LLMs for misinformation and what to do about that:

Generative Language Models and Automated Influence Operations: Emerging Threats and Potential Mitigations

• Generative language models have improved drastically, and can now produce realistic text outputs that are difficult to distinguish from human-written content. For malicious actors, these language models bring the promise of automating the creation of convincing and misleading text for use in influence operations. This report assesses how language models might change influence operations in the future, and what steps can be taken to mitigate this threat. We lay out possible changes to the actors, behaviors, and content of online influence operations, and provide a framework for stages of the language model-to-influence operations pipeline that mitigations could target (model construction, model access, content dissemination, and belief formation). While no reasonable mitigation can be expected to fully prevent the threat of AI-enabled influence operations, a combination of multiple mitigations may make an important difference.

Journalistic Lessons for the Algorithmic Age

• At The Markup we pioneered an array of scientifically inspired methods that used automation and computational power to supercharge our journalism. Reflecting on our work, I came up with 10 of the most important lessons I’ve learned using this approach.

Book

• Proofing chapters. Finished up to chapter 10. Minor tweaks

Brr.

SBIRs

• Meeting a 10:30 to discuss GPT with Isaac. Wide ranging and fun. He’s going to add some slides
• Afternoon chat with Aaron. Also wide ranging and fun. 1) We are probably in the Singularity, and 2) The universe is probably not a simulation
• After some struggling, got the dev branch of the binary encoding project set up with Rukan

Book

• Working on the proofs

GPT Agents

• The demo got accepted at IUI! I may be going to Australia?
• Getting the clustering and embedding working

Return glasses for less powerful prescription. I’ll do that after my 2:00 meeting

Looks like the end of academic access. Ah well, it was a nice run. Trained language models are more fun anyway

Extracting Training Data from Diffusion Models

• Image diffusion models such as DALL-E 2, Imagen, and Stable Diffusion have attracted significant attention due to their ability to generate high-quality synthetic images. In this work, we show that diffusion models memorize individual images from their training data and emit them at generation time. With a generate-and-filter pipeline, we extract over a thousand training examples from state-of-the-art models, ranging from photographs of individual people to trademarked company logos. We also train hundreds of diffusion models in various settings to analyze how different modeling and data decisions affect privacy. Overall, our results show that diffusion models are much less private than prior generative models such as GANs, and that mitigating these vulnerabilities may require new advances in privacy-preserving training.

And I found the Trump campaign trip I’ve been looking for!

SBIRs

• Finished the second draft! Need to send it out for some external sanity check. The SLT would like to see it too.
• 9:15 standup – done
• 11:30 CSC touch point
• 2:00 MORS meeting with Aaron – done! Sent off to SLT
• Send draft! Done!
• Check out GPT-Index (github.com/jerryjliu/gpt_index) – done! Need to see if it will work with Python 3.7.4
• Talk to Rukan and Aaron about making a separate repo for binary encoding project, notebooks, and results – done. Set up tomorrow maybe?

GPT-Agents

• Copy over and wire up PCA, TSNE, and DBSCAN.

Book

• Start proofing. I think downloading chapters to Word for grammar and spell checks is probably the way to go

This is true! I’ve put together a spreadsheet so you can see for yourself

SBIRs

• More FOM stuff. Maybe a meeting at 2:00?
• MORS paper with Aaron. Nope, but did finish the second draft.

GPT Agents

• 4:00 Meeting
• Went on a bit of a tangent discussing Bostrom’s paperclip conjecture and how recommender algorithms could be that, but from a human/ai source, not agi. The problem is at the scales that these systems might have effects at, it is not clear what the objective function means, and if we are, in fact destroying the world by creating an algorithm that seeks to optimize for one thing, but does so in ways that are ultimately destructive to humans. Venue could be the 5th AAAI/ACM Conference on AI, Ethics, and Society Papers are due on March 5.

Book

• The state of opinion toward gender, power, and policy

Tasks

• Glasses! Sent an email
• Physical! Message sent

Book

• Got the SkyLatex link for review. Need to be done by Feb 10

SBIRs

• 9:00 Sprint planning. Maybe get a chance to start with GPT-index? Done
• Continue with second draft
• Pre meeting with Aaron and Rukan
• Meeting with Loren. Since the goal is for each ship to “imagine” what the other ships are seeing and their FOM predictions, we need a way to have a way of easily positioning the ship position with respect to the threat in some shared, generalizable frame. And wrt the HGV, propagation seems… hard. Does it make more sense to simply simulate if an interception occurs at any time in the (recorded) flight paths? Then we can train the models on that.

GPT Agents

• Get embedding to work – done! Now I need to reduce and cluster
• IUI response delayed a few days

ChatGPT appears to be back! Tried asking it was to market the book. It came back with some good suggestions.

GPT Agents

• Continue working on automation
• Start on embedding?
• I think a larger test of shorter responses may be good for a proof-of-concept. About 64-128 tokens may be the sweet spot

SBIRs

• 9:00 Sprint demos
• 2:00 MDA weekly meeting

Book

• Based on the ChatGPT suggestions, I’m going to reach out to Wayne to see if he can connect me with a good reviewer. Maybe Ben Shneiderman? I could also try Roger, since he and his wife know everyone at UMD

Working on the NarrativeExplorer.

• Loading and saving of parameter files, so that it’s easy to store and try things from one session to the next
• Setting up automation
  • Need to strip things like multiple CRs and return them as a single. I’m getting one extra response. Fixed. I was calling the GPT twice. So it wasn’t one extra, it was double

One of Stacy’s friends had the really good ide about image and text generators could be a real boon for self held vision boards and planning. There has been a lot written about this, so they should be very good at handling the basic needs of folks

Overconfidently conspiratorial: Conspiracy believers are dispositionally overconfident and massively overestimate how much others agree with them – Gordon Pennycook, David G. Rand

• There is a pressing need to understand belief in false conspiracies. Past work has focused on the needs and motivations of conspiracy believers, as well as the role of overreliance on intuition. Here, we propose an alternative driver of belief in conspiracies: overconfidence. Across eight studies with 4,181 U.S. adults, conspiracy believers not only relied more intuition, but also overestimated their performance on numeracy and perception tests (i.e. were overconfident in their own abilities). This relationship with overconfidence was robust to controlling for analytic thinking, need for uniqueness, and narcissism, and was strongest for the most fringe conspiracies. We also found that conspiracy believers – particularly overconfident ones – massively overestimated (>4x) how much others agree with them: Although conspiracy beliefs were in the majority in only 12% of 150 conspiracies across three studies, conspiracy believers thought themselves to be in the majority 93% of the time.
• I think this could have an effect on stampede behavior more broadly. Something to the effect that when rulers (people with dominant power over others) are overconfident, they can more easily head in the direction of social realities (e.g. conspiracy theories, but also that VW could get away with cheating on emissions, or that the USA would not fail in Afghanistan).
• Overconfidence is a sort of dimension reduction since there is no need to look for complicated, nuanced positions. The most emotionally attractive answer is selected for and concentrates the overconfident.
• An implication for diversity injection is that the “landing page” for diversity has to be simple and emotionally attractive.

Science has finally cracked the mystery of why so many people believe in conspiracy theories (Business Insider article on the above)

Tasks

• Schedule physical
• See if my glasses are ready
• Chores

GPT Agents

• Wire up the loading of generator and embedding params. Maybe while I’m at it, read in a file with prompts and params? Done!
• Had a thought that rather than clustering, I could just work on distances and the number of connections at that distance. Too many connections is a node like “the”, and nodes with only two connections (the predecessor and successor in the narrative) may not be that interesting and could be discarded. Something to think about.
• Continue reading GPT-index documentation
• How Cohere Works with Google’s Vertex Machine Engine to Power Embeddings
  • We’ve put together a notebook on GitHub to help you learn how to create embeddings with the Cohere API and then leverage the Vertex AI Matching Engine to create and query an index. The notebook includes code samples and step-by-step instructions for using the Cohere Embed endpoint to quickly capture semantic information about input data, and then applying the Vertex AI Matching Engine’s Approximate Nearest Neighbor (ANN) service to find similar texts.

SBIRs

• Send Lauren the ppt to pass on
• 11:00 presentation – DONE!
  • Re-read the rationale from the paper
  • Have the paper open in as a PDF
• More editing and tweaking

GPT Agents

• Need to create a table for generator params and embedding params – Done! Hooked them up to the table_run and am saving them out on a per run basis. Next is to load them in

SBIRs

• Did a first pass at the ChatGPT slide deck
• Need to check out Lauren’s slides. Done – looks good. Still waiting for HGV video
• 9:15 standup
• Do a full read through and tweak of paper, then ping Angela and Paul about if they would be interested in reading. Pinged, but only got to page 10 or so. A good deal of rewriting and cleanup.
• Help Aaron with CONOPS paper? Done
• Finished up the slide deck and integrated Loren’s content.

GPT Agents

• Added a log to the social science and NNM sections
• Spent a good deal of time getting NarrativeExplorer built

• 4:00 Meeting

SBIRs

• Meeting with Lauren to go over slides. He’s going to add a few
• Postponed the commercialization meeting
• Need more FOM data. Need to talk to Rukan
• Tweaked War Elephants and added in some text about antifragility.

Book

• Got notification that the permissions team got my log. Hopefully they will be ok with it.

Nice intro to word and sentence embeddings from co:here – What Are Word and Sentence Embeddings?

Introduction to pynytimes – The New York Times is one of the most trusted news source around the world. All their article metadata is easily available using their API, which is publicly available to everyone (though only for non-commercial use). All this data can be queried using a REST API, however setting it up can be quite time-consuming. This library solves that problem, now you can easily and quickly query the API without having to worry about the specific implementation.

The techniques behind ChatGPT: RLHF, IFT, CoT, Read teaming, and more

• A few weeks ago, ChatGPT emerged and launched the public discourse into a set of obscure acronyms: RLHF, SFT, IFT, CoT, and more, all attributed to the success of ChatGPT. What are these obscure acronyms and why are they so important? We surveyed all the important papers on these topics to categorize these works, summarize takeaways from what has been done, and share what remains to be shown.

SBIRs

• 9:15 Standup
• 10:00 Q3 Slides meeting with Loren
• 1:00 Bi-weekly
• Get any responses back on paper (HA!) and get ready to send out

GPT Agents

• Set up schema. I’m thinking four tables: 1) Experiment (name, date, user, run number), 2) Experiment params 3) Text (text, embedding, projection, cluster ID) 4) Cluster (experiment, cluster_number, cluster_name, include/exclude) – done
• Add automation fields and buttons – done
• For development, load result text automatically – done
• Hooking up DB to App. Got a lot done. Experiments, runs, and text are stored using test data.

Dissociating language and thought in large language models: a cognitive perspective

• Today’s large language models (LLMs) routinely generate coherent, grammatical and seemingly meaningful paragraphs of text. This achievement has led to speculation that these networks are — or will soon become — “thinking machines”, capable of performing tasks that require abstract knowledge and reasoning. Here, we review the capabilities of LLMs by considering their performance on two different aspects of language use: ‘formal linguistic competence’, which includes knowledge of rules and patterns of a given language, and ‘functional linguistic competence’, a host of cognitive abilities required for language understanding and use in the real world. Drawing on evidence from cognitive neuroscience, we show that formal competence in humans relies on specialized language processing mechanisms, whereas functional competence recruits multiple extralinguistic capacities that comprise human thought, such as formal reasoning, world knowledge, situation modeling, and social cognition. In line with this distinction, LLMs show impressive (although imperfect) performance on tasks requiring formal linguistic competence, but fail on many tests requiring functional competence. Based on this evidence, we argue that (1) contemporary LLMs should be taken seriously as models of formal linguistic skills; (2) models that master real-life language use would need to incorporate or develop not only a core language module, but also multiple non-language-specific cognitive capacities required for modeling thought. Overall, a distinction between formal and functional linguistic competence helps clarify the discourse surrounding LLMs’ potential and provides a path toward building models that understand and use language in human-like ways.

Starting to read the documentation for GPT Index. It looks very thorough and capable. I need to get a charge number so I can dig into it and get paid.

SBIRs

• Working on the slide deck
• Contract stuff

GPT Agents

• Got the parsing done. Need to work on saving them to the deb and getting the embeddings. Also, I’ll need to set up a looping system that runs the prompt a specific number of times and does the parsing and storing. Something like “automate” with a field for how many times.

I loaded journal entries from the past 10 years into GPT-3—and started asking it questions – This might be the way to do my next online resume and the book website. Basically, index and summarize, which sounds perfect.

It uses GPT Index, a project consisting of a set of data structures designed to make it easier to use large external knowledge bases with LLMs. Lots of documentation here.

I realize that this approach could replace the finetuning of the GPT-2 models. This makes it extremely general. There could be a book agent, Twitter agent, or even a SBIR agent.

And because it is finding text “chunks” based on embedding similarity to the prompt, it can point back to the sources. That makes research into a subject where the corpora exists much better.

Going to dig into this some more.

Elicit.org is also doing something like this. Here’s the thread:

This is very good: