SBIRs

• Just grinding away on the proposal – finished something close to the final. Not bad for one week

Move some $$ around for contracting. May have to fire up the home equity LoC.

Testing and Evaluation of Health Care Applications of Large Language Models
A Systematic Review

• Of 519 studies reviewed, published between January 1, 2022, and February 19, 2024, only 5% used real patient care data for LLM evaluation. The most common health care tasks were assessing medical knowledge such as answering medical licensing examination questions (44.5%) and making diagnoses (19.5%). Administrative tasks such as assigning billing codes (0.2%) and writing prescriptions (0.2%) were less studied. For NLP and NLU tasks, most studies focused on question answering (84.2%), while tasks such as summarization (8.9%) and conversational dialogue (3.3%) were infrequent. Almost all studies (95.4%) used accuracy as the primary dimension of evaluation; fairness, bias, and toxicity (15.8%), deployment considerations (4.6%), and calibration and uncertainty (1.2%) were infrequently measured. Finally, in terms of medical specialty area, most studies were in generic health care applications (25.6%), internal medicine (16.4%), surgery (11.4%), and ophthalmology (6.9%), with nuclear medicine (0.6%), physical medicine (0.4%), and medical genetics (0.2%) being the least represented.

SBIRs

• 9:00 standup
• Work on proposal. I think finish up Technical, and start to figure out the SOW – done with the first draft of both! Tomorrow is the Quad chart

SBIRs

• Work on proposal. SOW and Quad chart, plus some more on the technical section. Got more done on technical, pulled out D2A because it was too much and not related. Changed CwoC to be more about using NNs to understand communication effectiveness WRT bandwidth and latency
• 12:50 USNA meeting. They showed their poster, which was good looking, but didn’t make that much sense
• 3:00 Tradeshow demo. Coming along. Nice box! I’ll need to be able to ssh into it to develop on.

There is a blissful lack of grim news to wake up to. This is about as big as it gets at 6:30AM EST:

Chores

• Clean house – done
• Bills – done
• Dishes – done
• Lawn = done
• Pick up truck, probably – needs more work
• Also, I welded a thing and remembered to turn off the nitrogen!

SBIRs

• Need to spend 2 hours on the proposal – wound up spending much more time on this because for some reason the technical section needed to be done today. So dumb

GPT Agents

• 4:15 Alden meeting – went well, though I swear we spent too much time on improving a strawman. Verified that you can get logprobs out of the legacy complete API

SBIRs

• 9:00 Standup
• 11:30 Catch up with Orest
• 4:30 book club
• Working a lot on the proposal first pass. Doing NNMs now. Got everything but BH/WH/AI done. Thinking about using the term “weapons-grade AI,” since weaponized is an overused term that has lost impact

GPT Agents

• 2:45 LLM meeting – we’re sending the paper in for an initial reaction

Brrr! Fall is here!

SBIRs

• Starting to work on the proposal. I want to use the concept of cyberspace, but as understood though embeddings. The hook is that the term “cyberspace” came out too early. It represented a need on the part of people experiencing the internet to be able to “navigate,” as opposed to “search.” It turned out that search was an easier problem, so we now have a search-based methodology for finding new content. Even recommender algorithms are search – they just use latent terms to promote items that will get your attention.
• But with “embeddings”, came a way to comprehend the vast amounts of online information in a spatial sense. The discovery that embeddings in deep neural network Language Models have a spatial relationship to one another that reflects human understanding is profound. That the equation king – man + woman = queen works in embedding space and matches our intuitive understanding implies that these models, trained on vast amounts of human-generated text, represent human understanding of information, belief, and opinion in discoverable ways.
• Games are at their core a way of exploring a domain constrained by rules in search of a winning condition. Neural networks and embeddings provide a new way to quantify that process, increase understanding, and increase the likelihood that novel winning solutions will not be overlooked.

• Made a lot of progress. Let’s see how tomorrow goes. Book club!

Nice break:

Breaking News in the Hands of a Few: Newsbrokering on X During the Trump Assassination Attempts

• A small number of accounts were highly prominent in the discourse on X/Twitter surrounding both the July and September 2024 assassination attempts on Donald Trump.
• We conceptualize the behaviors of these accounts as newsbrokering — the selective curation and dissemination of information by news influencers during breaking news events.
• Five of the nine most prominent newsbrokering accounts had previously been suspended from X or other platforms. 
• Traditional news outlets, despite having significantly more followers and twice as many tweets, struggled to compete with newsbrokering accounts in terms of audience engagement. 
• In both events, newsbrokering accounts not only curated and disseminated information but also often framed it to fit existing narratives and conspiratorial themes surrounding the assassination attempts. 
• Social media exhibits a trend toward the oligarchization of the news environment, where a few accounts can dominate important discourse. 

SBIRs

• 9:00 Standup
• 11:00 BAA meeting – checked the Overleaf – no new content. Looks like the CwoC piece may not make sense, but the NNM approach might. I also wonder if we could also some D2A where the current state of the game/sim is used as the configuration file and the likely outcomes are instantaneously calculated.

GPT Agents

• Found this nice quote in an article on the 2024 Nobel prize for Economics:
• “Countries with “inclusive” institutions that protected personal property rights and allowed for widespread economic participation tended to end up on a pathway to longer-term prosperity. Those that had what the researchers called “extractive” institutions — ones that helped elites to maintain control, but which gave workers little hope of sharing in the wealth — merely provided short-term gains for the people in power.”

Looks like Florida is getting hammered

Differential Transformer

• Transformer tends to overallocate attention to irrelevant context. In this work, we introduce Diff Transformer, which amplifies attention to the relevant context while canceling noise. Specifically, the differential attention mechanism calculates attention scores as the difference between two separate softmax attention maps. The subtraction cancels noise, promoting the emergence of sparse attention patterns. Experimental results on language modeling show that Diff Transformer outperforms Transformer in various settings of scaling up model size and training tokens. More intriguingly, it offers notable advantages in practical applications, such as long-context modeling, key information retrieval, hallucination mitigation, in-context learning, and reduction of activation outliers. By being less distracted by irrelevant context, Diff Transformer can mitigate hallucination in question answering and text summarization. For in-context learning, Diff Transformer not only enhances accuracy but is also more robust to order permutation, which was considered as a chronic robustness issue. The results position Diff Transformer as a highly effective and promising architecture to advance large language models.
• It’s interesting that “hallucinations,” which are interpolations between explicitly trained points may be a function of noise. This could be tuned to adjust the amount of “artistic license” in a model. Extremely noisy models may be the most interesting.

SBIRs

• 9:00 standup. Also, do slides for Monday and put in a story for this new white paper
• 9:30 Pre-proposal launch meeting
• Do the lunchtime ride around 11:00
• 12:50 USNA
• 2:00 Gigantor
• 4:30 Book club

GPT Agents

• 2:45 LLM Meeting. Should have lots to talk about

Bad day for hacked things:

Hacked ‘AI Girlfriend’ Data Shows Prompts Describing Child Sexual Abuse

• “It’s basically a handful of open-source projects duct-taped together. I started poking around and found some vulnerabilities relatively quickly. At the start it was mostly just curiosity but I decided to contact you once I saw what was in the database.”

Jim Donnies! Done

SBIRs

• 11:00 MP+SimAccel proposal meeting – went well, I think. Very different approaches. We’re new, bespoke, and they are legacy
• 1:30 LM MP tagup
• Work on book – working!

GPT Agents

• Ping everyone to say I’ve finished my pass through the paper
• add \t to bib script

SBIRs

• 9:00 Standup – done!
• LM White Paper? This can wait, actually. Ron will need to do the MLOps, and can’t even get started until returns from his BD tasks and trials

GPT Agents

• Continue with paper – done!
• Ping Greg to discuss his comments

Helene response hampered by misinformation, conspiracy theories

• Officials have sought to tamp down the misinformation that has continued to spread online. The Federal Emergency Management Agency has been updating a webpage seeking to dispute common rumors, while the North Carolina Department of Public Safety has done the same, writing that authorities were “working around-the-clock to save lives and provide humanitarian relief.”

AI-Generated Pro-North Korean TikToks Are Also Bizarre Ads for Supplements

• The ads also use an interesting blend of AI-generated cover images and real images within the slideshow itself. And not all of the ads are about North Korea. Some of them use AI-generated images of Taylor Swift and Jennifer Aniston to shill the supplements, while other slideshows are spreading disinformation about Mpox, are about TikTok trends like “demure,” or claim the supplements are “better than Ozempic.” 

Sample ballot of Baltimore County

Grants

• Finish review 14 – Done and submitted!

SBIRs

• Work on LM white paper with Aaron?
• 3:00 Demo kickoff meeting – mostly figuring out what resources (compute, screens, etc) will be needed

GPT Agents

• Work on paper. Wrote the script to convert footnotes to citations. It works well! Had a few issues getting raw strings to behave:

from tkinter import filedialog
import re
from typing import List, Dict

def load_file_to_list(filename:str) -> List:
    print("opening {}".format(filename))
    try:
        with open(filename, 'r') as file:
            lines = file.readlines()
            return [line.strip() for line in lines]
    except FileNotFoundError:
        print("Error: File '{}' not found".format(filename))
        return []

def save_list_to_file(l:List, filename:str):
    print("opening {}".format(filename))
    s:str
    try:
        with open(filename, 'w') as file:
            for s in l:
                file.write("{}\n".format(s))
    except FileNotFoundError:
        print("Error: File '{}' not found".format(filename))
        return []

filename = filedialog.askopenfilename(filetypes=(("tex files", "*.tex"),), title="Load tex File")
if filename:
    filename2 = filename.replace(".tex", "_mod.tex")
    filename3 = filename.replace(".tex", "_mod.bib")
    # open the pdf file
    l:List = load_file_to_list(filename)
    p1 = re.compile(r"\\footnote{\\url{(.*?)}}")
    p2 = re.compile(r"https://([^/]*)")

    s1:str
    s2:str
    s3:str
    l2 = []
    cite_dict = {}
    count = 1
    for s1 in l: # Get each line in the file
        #print(s1)
        m1 = p1.findall(s1) # find all the footnote urls
        for s2 in m1:
            #print("\t{}".format(s2))
            m2 = p2.match(s2) # pull out what we'll use for our cite
            s3 = m2.group(1).strip('www.')
            s3 = "{}_{}".format(s3, count)
            #print("\t\t{}".format(s3))
            olds = r"\footnote{\url{"+s2+"}}"
            news = r"\cite{"+s3+"}"
            #print("olds = {}[{}], news = {}".format(olds, s1.find(olds), news))
            s1 = s1.replace(olds, news)
            cite_dict[s3] = s2
        l2.append(s1)
        print(s1)
    save_list_to_file(l2, filename2) # write the modified text to a new file

    l2 = []
    for key, val in cite_dict.items():
        s = "@misc{"+key+",\n"
        s += '\tauthor = "{Last, First}",\n'
        s += '\tyear = "2024",\n'
        s += '\thowpublished = "\\url{'+val+'}",\n'
        s += 'note = "[Online; accessed 07-October-2024]"\n}\n'
        print(s)
        l2.append(s)
    save_list_to_file(l2, filename3) # write the citation text to a .bib file

U.S. Wiretap Systems Targeted in China-Linked Hack

• A cyberattack tied to the Chinese government penetrated the networks of a swath of U.S. broadband providers, potentially accessing information from systems the federal government uses for court-authorized network wiretapping requests.
• Here are my AI-weapons thoughts on this: 1) If you can plant a MitM LLM that works to make people want to legislate back doors for cybercrime, you could set up this kind of operation. 2) If these backdoors already exist, you can plant LLMs and cause further havoc, or adjust the behavior of your adversary in more subtle ways.

People Are Sharing Fake Hurricane Helene Photos for Profit and Political Gain

• “[Community Note] says this is AI. In this case, I don’t care. We should look out for our own (Americans before the rest of the world and I wouldn’t be at all surprised if there was a girl fitting the description that wasn’t lucky enough to make it to a photographer for such an image,” wrote yet another.

Where Facebook’s AI Slop Comes From

• Facebook itself is paying creators in India, Vietnam, and the Philippines for bizarre AI spam that they are learning to make from YouTube influencers and guides sold on Telegram.

An Audacious Plan to Halt the Internet’s Enshittification and Throw It Into Reverse

TPI-LLM: Serving 70B-scale LLMs Efficiently on Low-resource Edge Devices

• Large model inference is shifting from cloud to edge due to concerns about the privacy of user interaction data. However, edge devices often struggle with limited computing power, memory, and bandwidth, requiring collaboration across multiple devices to run and speed up LLM inference. Pipeline parallelism, the mainstream solution, is inefficient for single-user scenarios, while tensor parallelism struggles with frequent communications. In this paper, we argue that tensor parallelism can be more effective than pipeline on low-resource devices, and present a compute- and memory-efficient tensor parallel inference system, named TPI-LLM, to serve 70B-scale models. TPI-LLM keeps sensitive raw data local in the users’ devices and introduces a sliding window memory scheduler to dynamically manage layer weights during inference, with disk I/O latency overlapped with the computation and communication. This allows larger models to run smoothly on memory-limited devices. We analyze the communication bottleneck and find that link latency, not bandwidth, emerges as the main issue, so a star-based allreduce algorithm is implemented. Through extensive experiments on both emulated and real testbeds, TPI-LLM demonstrated over 80% less time-to-first-token and token latency compared to Accelerate, and over 90% compared to Transformers and Galaxy, while cutting the peak memory footprint of Llama 2-70B by 90%, requiring only 3.1 GB of memory for 70B-scale models.

SBIRs

• Look for new BAAs – done
• 9:00 Standup – done
• 12:45 USNA – trainwreck, hoping to fix
• Expense travel – done

GPT Agents

• Working on conclusions. I think I have a nice hook from the Gutenberg Parentheses
• 9:30 meeting with Matt. Show the GPM demo and talk about coordination with expensive information, stiff & dense networks vs. slack and sparse, the need for embodiment to find truly novel things, the curse of dimensionality, explore/exploit and the need for diversity. done!
• Respond to IUI – done

Grants

• Start review 14 – started