Tag Archives: Machine Learning

Phil 2.6.2025

From The Bulwark. Good example of creating a social reality and using it for an organizational lobotomy. Add to the book following Jan 6 section?

Full thread here

There is an interesting blog post (and thread) from Tim Kellogg that says this:

  • Context: When an LLM “thinks” at inference time, it puts it’s thoughts inside <think> and </think> XML tags. Once it gets past the end tag the model is taught to change voice into a confident and authoritative tone for the final answer.
  • In s1, when the LLM tries to stop thinking with “”, they force it to keep going by replacing it with “Wait”. It’ll then begin to second guess and double check it’s answer. They do this to trim or extend thinking time (trimming is just abruptly inserting “/think>”).

This is the paper: s1: Simple test-time scaling

  • Test-time scaling is a promising new approach to language modeling that uses extra test-time compute to improve performance. Recently, OpenAI’s o1 model showed this capability but did not publicly share its methodology, leading to many replication efforts. We seek the simplest approach to achieve test-time scaling and strong reasoning performance. First, we curate a small dataset s1K of 1,000 questions paired with reasoning traces relying on three criteria we validate through ablations: difficulty, diversity, and quality. Second, we develop budget forcing to control test-time compute by forcefully terminating the model’s thinking process or lengthening it by appending “Wait” multiple times to the model’s generation when it tries to end. This can lead the model to double-check its answer, often fixing incorrect reasoning steps. After supervised finetuning the Qwen2.5-32B-Instruct language model on s1K and equipping it with budget forcing, our model s1-32B exceeds o1-preview on competition math questions by up to 27% (MATH and AIME24). Further, scaling s1-32B with budget forcing allows extrapolating beyond its performance without test-time intervention: from 50% to 57% on AIME24. Our model, data, and code are open-source at this https URL

Tasks

SBIRs

  • 9:00 standup – done
  • 10:00 MLOPS whitepaper review
  • 12:50 USNA

Phil 9.27.2024

That is a lot of rain:

Learned feature representations are biased by complexity, learning order, position, and more

  • Representation learning, and interpreting learned representations, are key areas of focus in machine learning and neuroscience. Both fields generally use representations as a means to understand or improve a system’s computations. In this work, however, we explore surprising dissociations between representation and computation that may pose challenges for such efforts. We create datasets in which we attempt to match the computational role that different features play, while manipulating other properties of the features or the data. We train various deep learning architectures to compute these multiple abstract features about their inputs. We find that their learned feature representations are systematically biased towards representing some features more strongly than others, depending upon extraneous properties such as feature complexity, the order in which features are learned, and the distribution of features over the inputs. For example, features that are simpler to compute or learned first tend to be represented more strongly and densely than features that are more complex or learned later, even if all features are learned equally well. We also explore how these biases are affected by architectures, optimizers, and training regimes (e.g., in transformers, features decoded earlier in the output sequence also tend to be represented more strongly). Our results help to characterize the inductive biases of gradient-based representation learning. We then illustrate the downstream effects of these biases on various commonly-used methods for analyzing or intervening on representations. These results highlight a key challenge for interpretability—or for comparing the representations of models and brains—disentangling extraneous biases from the computationally important aspects of a system’s internal representations.

More AI slop:

Amazing to watch Google destroy its core functionality chasing AI. Friends on the groupchat were talking about Rickey Henderson, who threw left and hit from the right side, which is really rare. If you go to google to find other throw left/bat right players. This is what its AI gives you.

— Chris Hayes (@chrislhayes.bsky.social) 2024-09-27T18:49:17.788Z

From https://bsky.app/profile/chrislhayes.bsky.social/post/3l55tbzk5ue2e. He continues: “This is is garbage! It’s worst than useless, it’s misleading! If you looked at it quickly you’d think Babe Ruth and Shohei also both threw left and batted right. Sure this is trivial stuff but the whole point is finding accurate information.“

Chores today

  • Clean House – done!
  • Recycle – Done
  • Bills (TRP!) – Done
  • Prep for Seagull! – Done enough.

Grants

  • Finish proposal 14 – Good proposal! Now I need to write the reviews

Phil 4.15.2024

Tax day!

Read Collective intelligence: A unifying concept for integrating biology across scales and substrates, which is wild, and feeds into the prompt-as-life concept I’ve been toying with. Among other things, it opens up experiments to show the level of self-organization available to prompts:

  • A central claim of the emerging field of diverse intelligence is that cognitive capacities (Box. 1) exist on a spectrum: that tools, concepts, and approaches from behavioral sciences can be productively applied to understand and control systems far beyond familiar animals with central nervous systems (without the necessity to attribute advanced, human-level metacognitive traits). 
  • Biological intelligent systems demonstrate increased ability to achieve their (collective) goals despite obstacles by integrating the individual competencies of their components (which can perform tasks in their own space without any inkling of the large-scale goals to which they contribute)
  • Thus, the physiological process that leads to the emergence of integrated collectives, which scientists and conspecifics recognize as discrete individuals is fundamentally dependent on the geometry of interactions (and signaling barriers) present during the early establishment of individuality and the setting of borders between Self and outside world (since every cell is some other cell’s adjacent neighbor).
  • However, the more interesting and fundamental issue is seen when considering just one cut: the cells on either side of the cut will create a head and tail respectively, but they were adjacent neighbors before the cut and located at the same positional information value. In other words, it is actually impossible for an anatomical decision like this to be made locally – the cells of the wound must coordinate with the remaining fragment to get information about where they are located, which way they are facing, and what other structures exist121,122, in order to make adaptive decisions about large-scale growth and form that enable regeneration of normal worms.
  • This recruitment of individuals to accomplish a high-level goal is seen in other collective systems like ant colonies152,153, which often call in helpers when a task is large. The ability to recruit participants to complete tasks may be a central competency of collective intelligence that works across scales, from cells to swarms of entire organisms7.
  • Cell and developmental biology offer very rich fodder for the emerging field of diverse intelligence: discovering a vast spectrum of problem-solving capacities in novel substrates and at unconventional spatiotemporal scales. Because of life’s multi-scale competency architecture, a fundamental aspect of intelligence is collective behavior: all intelligences appear to be made of parts, connected by mechanisms implementing policies that bind the competent components into a cooperative (and competitive6) computational medium that solves problems in new spaces and at higher scales.
  • Importantly, the definition of intelligence as the ability to reach the same endpoint despite internal or external changes emphasizes not only robustness (successful use of novel navigational policies to overcome perturbations) but also its failure modes. Numerous ways of targeting of its sensory, memory, decision-making, or other components can de-rail the performance of a collective intelligence, resulting in birth defects and malformations.
    • I think this is a really important way to probe and examine prompts and models. How well do they reach their goals when damaged, and how do they do it.
  • Cancer, a kind of dissociative identity disorder of the somatic collective intelligence109, limitations in regenerative ability, and many physiological disorders could all be advanced by techniques that exploit not just the low-level mechanisms, but also the higher-level decision-making of life16,17. 
  • Living matter is a kind of agential material with the ability to propagate information across scales – a phenomenon which has many implications for evolution9, and for bioengineering21.

Ordered The Sentient Cell: The Cellular Foundations of Consciousness

SBIRS

  • Write email summary of Friday’s meeting. Also find out who I send the MCMC description to. Done
  • Start slide deck for the 22nd – started! Using ContextExplorer which is really good for this sort of thing.
  • Submit paper – done
  • Gotta rewrite the final report in a way that “substantially revises” it. Sigh. Waiting for some direction from someone in authority.