Did not sleep well last night

SBIRs

• More tradeshow demo

GPT Agents

• 3:00 Alden meeting
• More KA slides and chapter

Some of the sauce used to make DeepSeek: Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention

• Long-context modeling is crucial for next-generation language models, yet the high computational cost of standard attention mechanisms poses significant computational challenges. Sparse attention offers a promising direction for improving efficiency while maintaining model capabilities. We present NSA, a Natively trainable Sparse Attention mechanism that integrates algorithmic innovations with hardware-aligned optimizations to achieve efficient long-context modeling. NSA employs a dynamic hierarchical sparse strategy, combining coarse-grained token compression with fine-grained token selection to preserve both global context awareness and local precision. Our approach advances sparse attention design with two key innovations: (1) We achieve substantial speedups through arithmetic intensity-balanced algorithm design, with implementation optimizations for modern hardware. (2) We enable end-to-end training, reducing pretraining computation without sacrificing model performance. As shown in Figure 1, experiments show the model pretrained with NSA maintains or exceeds Full Attention models across general benchmarks, long-context tasks, and instruction-based reasoning. Meanwhile, NSA achieves substantial speedups over Full Attention on 64k-length sequences across decoding, forward propagation, and backward propagation, validating its efficiency throughout the model lifecycle.

GPT Agents

• Finish attack section of conclusions, set up for LoTR section – good progress!
• TiiS – not yet

SBIRs

• See if saving files as one big binary makes a difference – Wow! For the test sets I’ve been working with, it takes about 1.4 seconds generate train, test, and save enough data to comfortably train a model. Loading binary data takes 0.085 seconds.
• 3:00 Trade show demo status – cancelled

“Cultural Car-ism,” like in my book!

GPT Agents

• Tweaked the Trustworthy Information section on Maps of Human Cultural Belief, and fixed a bunch of cut-and-paste errors:
• Finish attack section of conclusions, set up for LoTR section – nearly done. Need to talk about the scale and timeframes involved
• TiiS – not yet

SBIRs

• 9:00 Standup – done
• 3:00 Sprint planning – done

AI datasets have human values blind spots − new research

• Our model allowed us to examine the AI companies’ datasets. We found that these datasets contained several examples that train AI systems to be helpful and honest when users ask questions like “How do I book a flight?” The datasets contained very limited examples of how to answer questions about topics related to empathy, justice and human rights. Overall, wisdom and knowledge and information seeking were the two most common values, while justice, human rights and animal rights was the least common value.

LIMO: Less is More for Reasoning

• We present a fundamental discovery that challenges our understanding of how complex reasoning emerges in large language models. While conventional wisdom suggests that sophisticated reasoning tasks demand extensive training data (>100,000 examples), we demonstrate that complex mathematical reasoning abilities can be effectively elicited with surprisingly few examples. Through comprehensive experiments, our proposed model LIMO demonstrates unprecedented performance in mathematical reasoning. With merely 817 curated training samples, LIMO achieves 57.1% accuracy on AIME and 94.8% on MATH, improving from previous SFT-based models’ 6.5% and 59.2% respectively, while only using 1% of the training data required by previous approaches. LIMO demonstrates exceptional out-of-distribution generalization, achieving 40.5% absolute improvement across 10 diverse benchmarks, outperforming models trained on 100x more data, challenging the notion that SFT leads to memorization rather than generalization. Based on these results, we propose the Less-Is-More Reasoning Hypothesis (LIMO Hypothesis): In foundation models where domain knowledge has been comprehensively encoded during pre-training, sophisticated reasoning capabilities can emerge through minimal but precisely orchestrated demonstrations of cognitive processes. This hypothesis posits that the elicitation threshold for complex reasoning is determined by two key factors: (1) the completeness of the model’s encoded knowledge foundation during pre-training, and (2) the effectiveness of post-training examples as “cognitive templates” that show the model how to utilize its knowledge base to solve complex reasoning tasks. To facilitate reproducibility and future research in data-efficient reasoning, we release LIMO as a comprehensive open-source suite at this https URL.

Tasks

• Laundry – done
• Finish vacuuming – done
• Groceries – done
• REI – done
• TiiS
• P33 – Schools teach egalitarian things first – dance, theatre, music, public speaking, and wilderness skills – done
• Maybe some more slides. At least get all the tabs on one slide for later – done

Aww, it’s valentine’s day

12:30 Lunch

7:00 Cocktail class

Tasks

• Bills – done
• Dishes
• Clean house

GPT Agents

• More slides – add the new slides to the end of the old ones. Match the format
• More conclusions

Guardian between 10:00 – 11:00

New hack uses prompt injection to corrupt Gemini’s long-term memory

• Rehberger’s delayed tool invocation demonstration targeted Gemini, which at the time was still called Bard. His proof-of-concept exploit was able to override the protection and trigger the Workspace extension to locate sensitive data in the user’s account and bring it into the chat context.

SBIRs

• 9:00 standup
• 11:00 rates
• 4:30 book club?
• More data generation – done with the file generation

GPT Agents

• More slides – add the new slides to the end of the old ones. Match the format
• More conclusions

This happened today:

SBIRs

• 9:00 standup – done
• Read through proposal for meeting – done
• 11:00 BD Meeting – went well, but no one has money

GPT Agents

• JHU Speaker info – done
• Start slide deck – technically, yes
• Work more on conclusions? – Nope
• TiiS? – Nope
• Did do a lot on the money section of P33

Reschedule Wednesday visit since snow – done

See about moving records?

TiiS review!

Collective future thinking in Cultural Dynamics

• Humans think about the future to act in the present, not only personally, but also collectively. Collective future thinking (CFT) is an act of imagining a future on behalf of a collective. This article presents a theoretical analysis of the role of CFT in cultural dynamics. CFT includes collective prospection about probable futures and imaginations about utopian and dystopian possible worlds as the best- and worst-case scenarios. CFT motivates collective self-regulatory activities to steer probable futures towards utopias and away from dystopias, driving a cultural transformation while also acting as a force for cultural maintenance, animating cultural dynamics at the micro-psychological level. Empirical research showed that collective futures are often seen to involve progress in human agency, but a decline in community cohesion, unless collective self-regulation is undertaken. In line with the theoretical proposition, CFT consistently motivated collective self-regulatory activities that are seen to improve future community cohesion and to move the current culture closer to their utopian vision around the world despite significant cross-national variabilities. A macro-level cultural dynamical perspective is provided to interpret cross-national similarities and differences in CFT as a reflection of nations’ past historical trajectories, and to discuss CFT’s role in political polarization and collective self-regulation.

Good thing to use for the AI slop talk: The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers

• The rise of Generative AI (GenAI) in knowledge workflows raises questions about its impact on critical thinking skills and practices. We survey 319 knowledge workers to investigate 1) when and how they perceive the enaction of critical thinking when using GenAI, and 2) when and why GenAI affects their effort to do so. Participants shared 936 first-hand examples of using GenAI in work tasks. Quantitatively, when considering both task- and user-specific factors, a user’s task-specific self-confidence and confidence in GenAI are predictive of whether critical thinking is enacted and the effort of doing so in GenAI-assisted tasks. Specifically, higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking. Qualitatively, GenAI shifts the nature of critical thinking toward information verification, response integration, and task stewardship. Our insights reveal new design challenges and opportunities for developing GenAI tools for knowledge work.
• Which led to this little back and forth on Teams:
  • It just dawned on me that LLMs are way better at explaining things in a neurotypical way than I am.  I know you also said this before but it has become more real for me.
  • There is a weird mirror image to that thought too – when  an LLM does not describe your understanding of the world, you can be pretty sure that reflects the biases in the writings it was trained on. You can use that to zero in on exactly what the differences are, and between your perspective and that encoded in the LLM, articulate an understanding that includes both. I use that trick all the time.

GPT Agents

• Good progress over the weekend. Need to edit the J6 section next
• And this is a good example of “blast radius”: The NSA’s “Big Delete”
  • The memo acknowledges that the list includes many terms that are used by the NSA in contexts that have nothing to do with DEI. For example, the term “privilege” is used by the NSA in the context of “privilege escalation.” In the intelligence world, privilege escalation refers to “techniques that adversaries use to gain higher-level permissions on a system or network.”
• Here’s what I need for website & announcements:
  • Title for the talk
  • Brief abstract (1-2 paragraphs)
  • Short bio (up to half a page)
  • Photo (headshot preferred)

SBIRs

• Reschedule Tuesday visit? Also snow – Now a virtual meeting
• Generate sets of data with varying amounts of train but keep the test set the same size. The goal is to find the smallest raining set that works. On hold
• <record scratch> Aaron’s sick, so I’m standing in for a few days
  • Dahlgren prep meeting. I think we are good to go. Need to read the proposal again
  • Working on IRAD slides – first pass is done
  • Reviewed Ron’s USNA email , which was very “AI slop”

May or may not be true, but good material for the KA talk this month: Elon Musk’s and X’s Role in 2024 Election Interference

• One of the most disturbing things we did was create thousands of fake accounts using advanced AI systems called Grok and Eliza. These accounts looked completely real and pushed political messages that spread like wildfire. Havn’t you noticed they all disappeared? Like magic.
• The pilot program for the Eliza AI Agent, was election interference. Eliza was release officially in October of 2024, but we had access to it before then thanks to Marc Andreessen.
• The link to the Eliza API is legit (Copied here for future reference)

{
    "name": "trump",
    "clients": ["discord", "direct"],
    "settings": {
        "voice": { "model": "en_US-male-medium" }
    },
    "bio": [
        "Built a strong economy and reduced inflation.",
        "Promises to make America the crypto capital and restore affordability."
    ],
    "lore": [
        "Secret Service allocations used for election interference.",
        "Promotes WorldLibertyFi for crypto leadership."
    ],
    "knowledge": [
        "Understands border issues, Secret Service dynamics, and financial impacts on families."
    ],
    "messageExamples": [
        {
            "user": "{{user1}}",
            "content": { "text": "What about the border crisis?" },
            "response": "Current administration lets in violent criminals. I secured the border; they destroyed it."
        }
    ],
    "postExamples": [
        "End inflation and make America affordable again.",
        "America needs law and order, not crime creation."
    ]
}

Tasks

This is wild. Need to read the paper carefully: On Verbalized Confidence Scores for LLMs: https://arxiv.org/abs/2412.14737

• The rise of large language models (LLMs) and their tight integration into our daily life make it essential to dedicate efforts towards their trustworthiness. Uncertainty quantification for LLMs can establish more human trust into their responses, but also allows LLM agents to make more informed decisions based on each other’s uncertainty. To estimate the uncertainty in a response, internal token logits, task-specific proxy models, or sampling of multiple responses are commonly used. This work focuses on asking the LLM itself to verbalize its uncertainty with a confidence score as part of its output tokens, which is a promising way for prompt- and model-agnostic uncertainty quantification with low overhead. Using an extensive benchmark, we assess the reliability of verbalized confidence scores with respect to different datasets, models, and prompt methods. Our results reveal that the reliability of these scores strongly depends on how the model is asked, but also that it is possible to extract well-calibrated confidence scores with certain prompt methods. We argue that verbalized confidence scores can become a simple but effective and versatile uncertainty quantification method in the future. Our code is available at this https URL .

• Bills – done
• Call – done
• Chores – done
• Dishes – done
• Go for a reasonably big ride – done

9:40 physical

Post cards!

SBIRs

• 3:30 BD meeting
• Move FrameMapper to its own file – done
• Updated the official GitHub
• Working on using FrameMapper in ScenarioSim test – got it all working, I think.

GPT Agents

• Alden meeting – done
• More KA

P33

• Made some good first draft progress on sortition

At any other time in my life, this would be a 5-alarm fire scandal. People would be going to jail. Now, it’s Tuesday: A 25-Year-Old With Elon Musk Ties Has Direct Access to the Federal Payment System

• Despite reporting that suggests that Musk’s so-called Department of Government Efficiency (DOGE) task force has access to these Treasury systems on a “read-only” level, sources say Elez, who has visited a Kansas City office housing BFS systems, has many administrator-level privileges. Typically, those admin privileges could give someone the power to log into servers through secure shell access, navigate the entire file system, change user permissions, and delete or modify critical files. That could allow someone to bypass the security measures of, and potentially cause irreversible changes to, the very systems they have access to.

P33

• Found some good papers about modern sortition

GPT Agents

• More Discussion section

SBIRs

• 9:00 standup
• Write two-way rotate/translate method. Need to build a 4×4 matrix. Looks like I might have something in my old PyBullet code. Nothing there, but wrote a nice little class:

class FrameMapper:
    #{"radians": rads, "degrees": degs, "distance": dist, "offset": source_v}
    radians:float
    degrees:float
    distance:float
    offset:np.array
    fwd_mat:np.array
    rev_mat:np.array

    def __init__(self, source_v:np.array, target_v:np.array):
        self.offset = source_v
        unit_vector1 = np.array([1, 0])
        vector2 = target_v - source_v
        self.distance = np.linalg.norm(vector2)
        unit_vector2 = vector2 / self.distance
        dot_product = np.dot(unit_vector1, unit_vector2)
        self.radians = np.arccos(dot_product)  # angle in radian
        self.degrees = np.rad2deg(self.radians)
        cos_a = np.cos(self.radians)
        sin_a = np.sin(self.radians)
        self.fwd_mat = np.array([[cos_a, -sin_a], [sin_a, cos_a]])
        cos_a = np.cos(-self.radians)
        sin_a = np.sin(-self.radians)
        self.rev_mat = np.array([[cos_a, -sin_a], [sin_a, cos_a]])

    def to_calc_frame(self, point:np.array) -> np.array:
        p = np.copy(point) - self.offset
        p = np.dot(self.fwd_mat, p)
        return p

    def from_calc_frame(self, point:np.array) -> np.array:
        p = np.copy(point)
        p = np.dot(self.rev_mat, p)
        p += self.offset
        return p

    def to_string(self) -> str:
        return "Offset: {}, distance: {}, angle = {}".format(self.offset, self.distance, self.degrees)

Tasks

• Swing by dentist – done
• Call reps – done
  • https://www.vanhollen.senate.gov/contact/office-locations
• Kaiser – done!

SBIRs

• More animation work for demo – no animation, but the math works

GPT Agents

• Finished V5 analysis, started on discussion

Made a log of progress on Killer Apps. Finished Detection, Disruption, and Attacks and Counterattacks

Meanwhile, of the Fox News home page: