Tasks

• International driver’s license
• Screen door
• Plants!
• Till

SBIRs

• 9:00 – Sprint Demos
• 12:30 Kickoff
• 3:00 Sprint Planning

Today on AI used with Ill intent:

• Baltimore County Police arrested Pikesville High School’s former athletic director Thursday morning and charged him with crimes related to the alleged use of artificial intelligence to impersonate Principal Eric Eiswert, leading the public to believe Eiswert made racist and antisemitic comments behind closed doors.

Also, it seems he probably scored high on the SDO scale: What it was like to be a student of Dazhon Darien, accused of framing principal with AI

• It was a good day for a few students at Pikesville High School on Thursday hanging out in the parking lot after school. Their former athletic director, who they said belittled them and made them feel uncomfortable, wasn’t coming back.

Can’t seem to backup my phone using itunes any more. Doing the cloud thing

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

• Humans are capable of strategically deceptive behavior: behaving helpfully in most situations, but then behaving very differently in order to pursue alternative objectives when given the opportunity. If an AI system learned such a deceptive strategy, could we detect it and remove it using current state-of-the-art safety training techniques? To study this question, we construct proof-of-concept examples of deceptive behavior in large language models (LLMs). For example, we train models that write secure code when the prompt states that the year is 2023, but insert exploitable code when the stated year is 2024. We find that such backdoor behavior can be made persistent, so that it is not removed by standard safety training techniques, including supervised fine-tuning, reinforcement learning, and adversarial training (eliciting unsafe behavior and then training to remove it). The backdoor behavior is most persistent in the largest models and in models trained to produce chain-of-thought reasoning about deceiving the training process, with the persistence remaining even when the chain-of-thought is distilled away. Furthermore, rather than removing backdoors, we find that adversarial training can teach models to better recognize their backdoor triggers, effectively hiding the unsafe behavior. Our results suggest that, once a model exhibits deceptive behavior, standard techniques could fail to remove such deception and create a false impression of safety.

Followup: Simple probes can catch sleeper agents

• This “Alignment Note” presents some early-stage research from the Anthropic Alignment Science team following up on our recent “Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training” paper. It should be treated as a work-in-progress update, and is intended for a more technical audience than our typical blog post. This research makes use of some simple interpretability techniques, and we expect to share more results from collaborations between our Alignment and Interpretability teams soon.

Related: Coup probes: Catching catastrophes with probes trained off-policy

• We present experiments measuring the generalization abilities of probes trained off-policy in a toy setting. We show that probes can generalize well to different text formats and also generalize from harmful text the LLM wouldn’t output to harmful text where the LLM has been jailbroken to actually output the harmful text.

Related: How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions

• Large language models (LLMs) can “lie”, which we define as outputting false statements despite “knowing” the truth in a demonstrable sense. LLMs might “lie”, for example, when instructed to output misinformation. Here, we develop a simple lie detector that requires neither access to the LLM’s activations (black-box) nor ground-truth knowledge of the fact in question. The detector works by asking a predefined set of unrelated follow-up questions after a suspected lie, and feeding the LLM’s yes/no answers into a logistic regression classifier. Despite its simplicity, this lie detector is highly accurate and surprisingly general. When trained on examples from a single setting — prompting GPT-3.5 to lie about factual questions — the detector generalises out-of-distribution to (1) other LLM architectures, (2) LLMs fine-tuned to lie, (3) sycophantic lies, and (4) lies emerging in real-life scenarios such as sales. These results indicate that LLMs have distinctive lie-related behavioural patterns, consistent across architectures and contexts, which could enable general-purpose lie detection.

SBIRs

• 9:00 Standup
• 3:00 AFRL meeting – looks like we’ll set up an overleaf project and start generating a white paper every few months. Topic 1-(something) will be first. Going to see what goes on with the MORS talk first?
• 4:00 ONR meeting – We can repurpose the M30 content into the slide format, then maybe do that with the AFRL white papers
• 4:30 Book club

GPT Agents

• 2:00 Meeting

Or 4/24/24. Or 24/4/24, which also looks nice

SBIRs

• 1:30: Some CwoC discussion? Yup
• Spent the rest of the day setting up my dev environment

Woke up nice and relaxed after a good night’s sleep. The night before a presentation is not easy for me.

I’ve been thinking about this slide from the talk yesterday:

I think that AI researchers are in a place that nuclear researchers were in the ’30’s. There is this amazing technology that is going to change the world, but no one is sure how. Then the world engages in a total war that depends on technology and the Allies are not doing well. Some of the researchers think that a nuclear weapon might turn the tide. It works, but in retrospect it was too much too late. But for 10 years the chance that there could be a broad nuclear war was high, and take as just an extension of current developments – a bigger bomb. It took decades for that viewpoint to shift. AI weapons are probably here already, and there are nations and organizations that are working out the best way to use them – as an extension of current “active measures” strategies and tactics. And like the atomic bomb, we really have no idea where this will go.

SBIRs

• Read a bunch of stuff for upcoming meetings
• Fire up the NNM instance and see if I can remember how to use it. Add an instruction section to the notebook – got sidetracked into doing a detailed read of a BAA
• 9:00 standup
• 11:30 AI Ethics training discussion with Hall Research. They are legit as it gets. Let’s see what kind of training they put together, but for now I give a ringing endorsement
• 3:30 meeting on the Phase IIe. We have three weeks to respond, but it doesn’t seem like they are asking for much? Very confused. Maybe because it’s an extension?

Finished the slide deck and gave the talk. A single question. Still, it’s a nice deck that could get used elsewhere. Also need to update my CV – done

SBIR’s

• Need to show Protima how to set up an Overleaf project for research documentation.
• Need to start again on the NNM code, but at 4:30, it’s too late in the day

Chores

• Ground rent research
• House cleaning
• Yard
• See if truck is ready – back in the driveway, and it’s electronics magically fixed themselves when it was getting its oil change at no charge. I’m agape.
• BSO
• Water bill

SBIRs

• Slides <- could have done more, but now I have a PLAN!

Dentist!

7:30 syphony

SBIRs

• Did some work on the slides but not enough.
• Helped Aaron on the paper and go motivated to create an ASRC template so we never have to do *that* again
• 9:00 standup
• 11:00 CUI meeting – cancelled
• 4:30 book club – cancelled for the best – and saddest – of reasons

GPT-Agents

• Set Alden’s overleaf project up. Probably overkill
• 2:00 meeting – Just Jimmy, but fun!
• CUI provocation v3 submitted

SBIRs

• Big kerfuffle on the report yesterday afternoon. As a result, I just worked on it till it was done. Submitted this morning to show my commitment. Sigh
• Need to work on the slide deck today but my motivation is lacking

World’s First AI Pageant To Judge Winner On Beauty And Social Media Clout

SBIRs

• Fold KillerApps slides into presentation, then start thinking about images
• 9:00 standup
• 2:30 AI ethics maybe?
• Need some guidance on the way to handle the “culling,” which I think is going to be pretty weird unless the writing around the culls are reworked.

Conquering the COVID-19 Infodemic: How the Digital Black Press Battled Racialized Misinformation in 2020

• In 2020, as many Black people around the world fought both anti-Black racism and COVID-19, the Black press in the US was dealing with another widespread problem: an infodemic. Editors of Black digital publications were on the frontlines of dispelling racialized misinformation about COVID-19, all while reporting on a contentious presidential election, ongoing protests for racial justice, and a rising COVID-19 death toll that disproportionately affected African Americans. This mixed-methods study—which includes semi-structured interviews in addition to website and social media analyses—explains the top five tactics that Black outlets used to serve as an advocate for, and an adviser to, their communities during its time of dire need. Their strategies provided an editorial slant that challenged anti-Black racism in public discourse and countered misinformation with factual public interest journalism.

SBIRs

• Driving 4 hours for a 2 hour meeting
• It went pretty well, but Dr. J was a no show. Everyone is very interested in MP-type stuff, so I think we should organize the DTA work to produce a demo along these lines, maybe within the context of the MDBE. That could look very fancy.
• Wrote up a easy to read Monte Carlo Markov Chain description.
• The big use for NNMs (that no one can grok) is the visualization and prediction of mismanagement. It’s all people talk about in these places.

Markov Chain Monte Carlo: Exploring Probability Through Random Walks

Imagine you’re trying to figure out how effective a new drug is at treating a disease. The traditional statistical methods might not work very well because the problem is too complex – there are just too many factors to consider. This is where Markov Chain Monte Carlo (MCMC) can really shine.

MCMC is a powerful technique that combines two key ideas: Markov chains and Monte Carlo simulations.

A Markov chain is a sequence of events where each step only depends on the previous one. It’s like a random walk, where your next move is based only on where you are now, not on your whole history.

Monte Carlo simulations, on the other hand, are all about playing with randomness to find answers. Instead of trying to calculate everything exactly, you take a bunch of random samples and use those to estimate what you want to know.

Put these two ideas together, and you get MCMC. The basic process is:

• Start with an initial guess about the drug’s effectiveness.
• Take a small, random step from that starting point. This step represents the uncertainty or noise in the data.
• Decide whether to keep this new guess based on how well it fits the data you have. The better it fits, the more likely you are to accept it.
• Repeat steps 2 and 3 many, many times.

Over time, as you keep taking these random steps and accepting or rejecting them, your guesses will start to converge towards a meaningful result. This convergence is key – it tells you that the MCMC process has thoroughly explored the probability distribution and found a stable estimate.

MCMC is really powerful because it can handle complex models and uncertainties that would be very difficult to deal with using traditional methods. In our drug example, MCMC could help you estimate the drug’s effectiveness while accounting for all sorts of factors like patient characteristics, side effects, and so on.

The great thing about MCMC is that it’s flexible and can be adapted to all kinds of research problems, from predicting disease progression to optimizing drug combinations for cancer treatment. By leveraging the power of random walks and probability, MCMC can turn complexity into clarity and uncertainty into insight. It’s a truly remarkable tool in the researcher’s toolkit.

Chores:

• Call Jim Donnie’s
• Schedule oil change

SBIRs

• Finishing and submitting CUI paper
• 4:30 Book club
• Installed an ancient printer driver and it worked!

GPT Agents

• 2:00 LLM Meeting

The eclipse was cool. Even with clouds, it’s magical:

That’s the view to where the sun is still shining. We could also see the eclipse faintly through the clouds, which must have been how most people saw it before rapid travel. The light was so faint, it was easy to imagine the sun and the moon fighting. I swear I saw sparks.

SBIRs

• Finish CUI paper (done!) and prep for submit (topics, keywords, etc)
• Add UMBC talk to stories – done
• Contact Ian with info – done. Looks like it is a 1 hour talk!
• Respond to Kate S.

Prepping to go to the Eclipse! The forecast is looking pretty good!

SBIRs

• Need to add a story on the April 22 presentation to the IS Dept. Also need to write up the abstract!
• Work on the CUI 2024 submission. Had some fun with MidJourney making a teaser image. I need to brush up on blending in Photoshop

GPT Agents

• Nice chat, went over the white hat AI concept some more. Maybe something for an NSF proposal?
• 4:00 Meeting at UMBC. I got to play wise sage. Fun!