Grigori Guitchounts
AI + biology, venture creation @FlagshipPioneer | Neuroscientist @Harvard | Writer, pianist, runner, painter, dilettante
- I keep coming back to this: one protein sequence doesn’t cash out to one behavior. Even “the same” molecule can hop between shapes, with different timing. And the weird, rare states—the ones you’d bet against—can still matter.
- Single-molecule tools are great for the clockwork: dwell times, transition rates, how one molecule changes over minutes. Sequencing / -omics are great for the census: which variants are in the jar, and roughly how many of each.
- The nasty part is stitching them together: taking a kinetic trace and saying, confidently, this trace came from *this exact* molecule (sequence, PTMs, etc.)—without smearing out the time resolution that makes single-molecule worth it.
-
View full threadFrom sequence to function: bridging single-molecule kinetics and molecular diversity (Science) science.org/doi/10.1126/science…
- Your blood is full of cell-free DNA—millions of tiny shards, like confetti after a rough party. An Alzheimer’s classifier trained on them ends up leaning hard on a blunt signal: fragment length patterns, not just sequence or methylation calls.
- What grabbed me was the workflow: train a big epigenetics foundation model, then actually ask, via interpretability, “what does the AD head *use*?” instead of politely nodding at a black box and moving on.
- When they poked the embeddings, fragment-level methylation + fragment length were easy to read out. Genomic locus/region identity was fuzzier. That smells like the representation is more “how the DNA broke” than “where it came from.”
-
View full threadInterpretability to surface a new Alzheimer’s biomarker class (fragmentomics): goodfire.ai/research/interpreta…
- People say “emergence” in LLMs like it’s a magic trick: nothing… nothing… then—poof—capability. In complexity science it’s stricter. Emergence is when you can describe the system in a new, lower-dimensional way that makes the messy micro-details irrelevant.
- This paper basically says: a kink in a benchmark curve isn’t enough. If you want to claim “emergence,” show (1) real task success and (2) evidence the network built a new coarse-grained representation—an effective variable/model that makes prediction simpler.
- A distinction that stuck with me: emergent capabilities vs emergent intelligence. Capabilities can be narrow and trained-in, like a calculator doing its little routines. Intelligence, here, is “more with less”: reusing the same coarse-grainings widely, often by analogy.
-
View full threadLarge Language Models and Emergence: A Complex Systems Perspective (Krakauer, Krakauer, Mitchell) arxiv.org/pdf/2506.11135
- KG retrieval has an irritating failure mode: either you cast a wide net (nice coverage, but it’s all a bit mushy) or you commit to edge-walking (great multi-hop… unless you picked the wrong starting node and everything collapses). Real queries usually want both, in one pass.
- @marinkazitnik’s ARK gives the model two explicit buttons to press—(1) global lexical search over node descriptions, and (2) one-hop neighborhood expansion. “Multi-hop” is just pressing the second button again and again.
- Once you can alternate, you’re not gambling on a single seed node up front. You also don’t have to pre-declare some hop budget like it’s a road trip itinerary. You can fan out, find the right neighborhood, then start following relations with purpose.
-
View full threadPaper: arxiv.org/abs/2601.13969
- LLMs can do a decent impression of almost anyone… until they can’t, and you feel the rubber band snap back to “Helpful Assistant.” This paper tries to locate that snap-back in the model’s activations—and finds what looks like a single direction for “Assistant-ness.”
- How they do it: prompt the model to roleplay hundreds of archetypes, save the activations, then build “role vectors” (activation differences between personas). Run PCA on those vectors and the top component keeps reappearing: an Assistant Axis across models.
- What caught my attention: you can steer along this axis and it mostly does what you’d expect. Push toward it and you get the familiar helpful/harmless assistant vibe. Push away and it starts “being” other entities more readily. Far enough and it gets… mystical, theatrical.
-
View full thread
- Hypotheses are getting cheap. Lab time isn’t. A lot of “AI for science” feels like moving the traffic jam: from dreaming up ideas to the grimy work of checking them—what you test, how quickly, and what you do when the first run faceplants.
- ARIA’s funded “AI Scientist” projects are built around that constraint. The pitch isn’t “can it write a convincing protocol?” It’s: can an AI actually run a physical loop—design → test → update—without falling apart the moment reality pushes back.
- I have a soft spot for the nine‑month sprint. It’s short enough to force something real onto the bench, and long enough to hit the unglamorous stuff—failed assays, drifting sensors, confounders that show up like lint, shipping delays.
-
View full threadIf these systems can’t answer “the experiment went sideways—now what?”, they won’t matter. If they can, the payoff might be a different kind of progress: not smarter hypotheses, but faster, steadier iteration in the physical world. aria.org.uk/ai-scientist/funded…