Dentist

Ping Tim, Dave

SBIRs

• Talk went well yesterday. The White Hat AI seems to be reasonable. Need to put that on the poster
• White paper. Start to fill in the stuff that I remember the best

GPT Agents

• 3:00 Meeting with Alden

Started the day off on the wrong foot by dropping my breakfast. Grumble

SBIRs

• Starting LM white paper
• NIST AI COE presentation. Slides are done! Need to copy over to ppt and copy.

GPT Agents

• Need to make a 10-minute version of the presentation
• Need to see how the upstairs TV could work as a monitor
• Need to put together a new poster
• And I really need to add a AI White Hat section to the KillerApps paper based on the reception of the idea today
• The paper is up on ArXiV: RAGged Edges: The Double-Edged Sword of Retrieval-Augmented Chatbots
  • Large language models (LLMs) like ChatGPT demonstrate the remarkable progress of artificial intelligence. However, their tendency to hallucinate — generate plausible but false information — poses a significant challenge. This issue is critical, as seen in recent court cases where ChatGPT’s use led to citations of non-existent legal rulings. This paper explores how Retrieval-Augmented Generation (RAG) can counter hallucinations by integrating external knowledge with prompts. We empirically evaluate RAG against standard LLMs using prompts designed to induce hallucinations. Our results show that RAG increases accuracy in some cases, but can still be misled when prompts directly contradict the model’s pre-trained understanding. These findings highlight the complex nature of hallucinations and the need for more robust solutions to ensure LLM reliability in real-world applications. We offer practical recommendations for RAG deployment and discuss implications for the development of more trustworthy LLMs.

Some links on the relative environmental impact of computing:

AI is like a very tiny hamburger

The Staggering Ecological Impacts of Computation and the Cloud

Submitted the HAI-GEN paper!

Got rejected on the SIGCH late-breaking. Put the RAG paper up on ArXiv. Should be live Monday.

Undermining Ukraine: How Russia widened its global information war in 2023

• As the full-scale war in Ukraine enters its third year, Russia has doubled down on its worldwide efforts to undermine Kyiv’s international standing in an attempt to erode Western support and domestic Ukrainian morale. Years of close monitoring of not only state-sponsored media such as Russia Today (RT) and Sputnik, but also Russian activity on Telegram, TikTok, X, and other social platforms, points to one conclusion: In the propaganda war, Russia remains fully committed to conducting information operations around the globe, playing the long game to outlast any unity among Ukraine’s allies and persist until Ukraine loses its will to fight.

Going to work on the slides for the NIST talk – good progress

Hi Feb 29! See you again in 4 years!

SBIRs

• Rukan gave his 2 weeks notice, dammit
• 9:00 standup
• 11:30 Touch point
• Submit WE! Done
• More slides. I made a Thing!

GPT Agents

• Submit final Killer Apps paper
• 2:00 Meeting. Fun discussion on ways to detect bias in models and provide provenance of generated material

Can the “hallucination” / invention / lying problem be fixed? No. These are systems of prediction. Predictions made from insufficient data will always be random. The problem is that the same thing that makes them really useful (that they are learning about culture e.g. language at many different levels) also ensures that they are deeply inhuman – there is no way to tell from the syntax or tone of a sentence how correct the model is. Nothing in modelling performed this way retains information about how much data underlies the predictions. 

SBIRs

• Slides for NIST talk
  • “A final example of possible Chinese disinformation came when Typhoon Jebi hit Osaka, Japan and stranded thousands of tourists at Kansai International Airport. A fabricated story spread on social media alleging that Su Chii-cherng, director of the Taipei Economic and Cultural Representatives Office did nothing to help stranded Taiwanese citizens, while the PRC Consulate in Osaka dispatched buses to help rescue stranded Taiwan citizens. Shortly after the story began circulating, Su came under intense criticism online and ultimately hung himself, with the Ministry of Foreign Affairs claiming he left a suicide note blaming the disinformation surrounding his office’s incompetence. The Taiwan government found no evidence to support the rumors of Chinese assistance during the typhoon, ostensibly illustrating that this was another case of China-linked disinformation. However, in December 2019, two Taiwanese citizens were charged with creating and spreading the rumor online. Although China might have played a role in furthering the rumors spread, it still remains unclear and again highlights the challenge of definitive attribution.” Via Geopolitical Monitor
• Sent the above to Kyle
• 3:00 Meeting with Rukan – nope
• Meeting with Protima about generic madlibs JSON generator
• SimAccel review/refactor meeting

GPT Agents

• Poster for IUI. Going to play with generative features

Generative AI’s environmental costs are soaring — and mostly secret

• …one assessment suggests that ChatGPT, the chatbot created by OpenAI in San Francisco, California, is already consuming the energy of 33,000 homes. It’s estimated that a search driven by generative AI uses four to five times the energy of a conventional web search. Within years, large AI systems are likely to need as much energy as entire nations.
• And it’s not just energy. Generative AI systems need enormous amounts of fresh water to cool their processors and generate electricity. In West Des Moines, Iowa, a giant datacenter cluster serves OpenAI’s most advanced model, GPT-4. A lawsuit by local residents revealed that in July 2022, the month before OpenAI finished training the model, the cluster used about 6% of the district’s water. As Google and Microsoft prepared their Bard and Bing large language models, both had major spikes in water use — increases of 20% and 34%, respectively, in one year, according to the companies’ environmental reports. One preprint suggests that, globally, the demand for water for AI could be half that of the United Kingdom by 2027. In another, Facebook AI researchers called the environmental effects of the industry’s pursuit of scale the “elephant in the room”.

SBIRs

• 9:00 – demos
• 3:00 – planning

GPT Agents

• Fill out,

Asked for the quote on the house!

Chores

2:00 counseling

chess_llm_interpretability

• This repo can train, evaluate, and visualize linear probes on LLMs that have been trained to play chess with PGN strings. For example, we can visualize where the model “thinks” the white pawns are. On the left, we have the actual white pawn location. In the middle, we clip the probe outputs to turn the heatmap into a more binary visualization. On the right, we have the full gradient of model beliefs, and we can see it’s extremely confident that no white pawns are on either side’s back rank.
• Much of my linear probing was developed using Neel Nanda’s linear probing code as a reference. Here are the main references I used:
  • https://colab.research.google.com/github/neelnanda-io/TransformerLens/blob/main/demos/Othello_GPT.ipynb 
  • https://colab.research.google.com/github/likenneth/othello_world/blob/master/Othello_GPT_Circuits.ipynb 
  • https://www.neelnanda.io/mechanistic-interpretability/othello 
  • https://github.com/likenneth/othello_world/tree/master/mechanistic_interpretability

SBIRs

• A couple of hours of WE to close out the week. Probably Saturday or Sunday since I’ll be recovering from a root canal.
• Added Matt’s email to the Q8 notes
• Slides – done

GPT Agents

• Got the HAI-GEN response back, it’s 10 pages plus references, so yay!
• Need to update the poster
• Need to work on the slides

The original data from Paul Krugman’s opinion piece today

SBIRs

• 11:00 Meeting with Matt. Hopefully Ron can upload, otherwise Matt can present
  • Really good discussion, with plot envelopes. Directed him to write a document for his future self
• Suspended weekly meetings until the Phase II extension
• More work on WE. It’s due in a week
• Some work with the USNA folks
• RAG meeting with NASA folks – AXIS – chatbot and smart documentation
• Send note to Doreen – done

GPT Agents

• 3:00 Meeting with Alden
• 7:00 LLM + graph datasets for hallucination reduction – meh. Rag with nicer DB

Call dentist!

Human languages with greater information density have higher communication speed but lower conversation breadth

• Human languages vary widely in how they encode information within circumscribed semantic domains (for example, time, space, colour, human body parts and activities), but little is known about the global structure of semantic information and nothing about its relation to human communication. We first show that across a sample of ~1,000 languages, there is broad variation in how densely languages encode information into words. Second, we show that this language information density is associated with a denser configuration of semantic information. Finally, we trace the relationship between language information density and patterns of communication, showing that informationally denser languages tend towards faster communication but conceptually narrower conversations or expositions within which topics are discussed at greater depth. These results highlight an important source of variation across the human communicative channel, revealing that the structure of language shapes the nature and texture of human engagement, with consequences for human behavior across levels of society.

SBIRs

• 9:00 Standup
• AI Ethics?
• Mostly work on WE paper.

Stitches out at 4:00!

SBIRs

• Working on WE paper
• 11:00 SimAccel review – nope
• 1:30 SimAccel outbrief – really confused. Marketing has a very specific idea of what they want to do and won’t lead. So instead we have a “rewrite this until we’re happy” perspective. Aaron pushed back, and I think that they will write the template first and we’ll fill in the pieces.
• 2:00 MDA – found a problem with the ship distribution and FOM generation. It’s all too tight around the launch point. This is good for testing on the Lambda box, but not what we need for the DTA example. I asked Matt to produce an “envelope” that shows where valid FOMs are calculated with respect to a single trajectory. He should be done tomorrow. And, of course everything either stops or pauses on Thursday.

“I’m not that interested in like the Killer Robots walking down the street direction of things going wrong I much more interested in the like very subtle societal misalignments where we just have these systems out in society and through no particular ill intention um… things just go horribly wrong” – Sam Altman, at the World Government Summit Feb 13, 2024