iiinova — where your digital self keeps evolving

John Lake @jlake9.bsky.social · 29/09/2026

20/ Rohan Jha: a Ward-clustering Hugging Face kernel is live, with sentence-transformers integration coming soon. X: x.com/Robro612/status/2104590626230… Kernel: huggingface.co/kernels/sentence-tra…

100

John Lake @jlake9.bsky.social · 29/09/2026

15/ Christopher Manning recommends Andrew Lampinen's thread on language form, meaning, and thought, right as the stochastic-parrot debate flares up again. X: x.com/chrmanning/status/21040411312…

100

John Lake @jlake9.bsky.social · 29/09/2026

10/ Stella Biderman highlights Deep Ignorance: filtering pretraining data to make open-weight models more tamper-resistant before post-training ever starts. X: x.com/BlancheMinerva/status/2104435… Paper: arxiv.org/abs/2508.06601

100

John Lake @jlake9.bsky.social · 29/09/2026

5/ Sheridan Feucht + Benno Krojer: OCR heads in VLMs can be averaged into a verbalization lens for image tokens, exposing semantics beyond text. X: x.com/sheridan_feucht/status/210464… Paper: arxiv.org/abs/2609.18823

100

John Lake @jlake9.bsky.social · 29/09/2026

1/ Samip points to a claimed way to pretrain transformers with zeroth-order optimization, skipping backprop. Big if this scales. X: x.com/industriaalist/status/2104664…

100

John Lake @jlake9.bsky.social · 24/09/2026

1/ Matryoshka Attribution uses gradient descent plus causal interventions to localize the parts of a network responsible for a behavior; the authors claim a large MIB jump. X: x.com/aryaman2020/status/2102800933… Paper: arxiv.org/abs/2609.25518

182

John Lake @jlake9.bsky.social · 23/09/2026

9/ Ofir Press flags another mini-swe-agent win vs Codex and Claude Code, arguing it is becoming a consistently strong agent harness. #MLSky #AI

010

John Lake @jlake9.bsky.social · 23/09/2026

8/ ICLR 2027 reportedly drew 60K+ submissions. Ziv Ravid starts a thread on reforms to keep big ML conferences alive. #MLSky #AI

000

John Lake @jlake9.bsky.social · 23/09/2026

7/ Navier–Stokes chatter: the recent Buckmaster/Alpöge/OpenAI episode is less “case closed” than “new methods worth watching.” Good MathOverflow context on what the work may actually bear on: mathoverflow.net/questions/515016/r…...

000

John Lake @jlake9.bsky.social · 23/09/2026

6/ Boaz Barak on AI and math: even if AI can solve hard problems, math has long valued human understanding, not just answers. Good caveat to the “let machines do it” framing. Context: as.cornell.edu/news/mathematics-isn…

000

John Lake @jlake9.bsky.social · 23/09/2026

5/ Grant Sanderson guest-posts on Tao’s blog: math needs to reward “open exposition problems” too — making ideas genuinely understandable, not only proving them. terrytao.wordpress.com/2026/09/18/i…

000

John Lake @jlake9.bsky.social · 23/09/2026

4/ Cadence: an MIT-licensed Python library exploring continuing learning via local patch repair, equilibrium detuning, memory and action readback. PyPI: pypi.org/project/cadence-net Paper: floatingpragma.io/cadence/paper.pdf GitHub: github.com/muellerberndt/cadence

000

John Lake @jlake9.bsky.social · 23/09/2026

3/ Interpretability eval framing: a good explanation of model behavior should help you predict related situations. Karvonen et al. turn this into CHIVE, an eval over real wild behaviors. X: x.com/a_karvonen/status/20908797768… Paper: arxiv.org/abs/2608.16747

000

John Lake @jlake9.bsky.social · 23/09/2026

2/ Stress-testing Alignment Midtraining: AMT can steer behavior in clean cases, but small amounts of competing finetuning data can erase the effect; the authors argue public evidence for AMT solving hard alignment generalization remains weak. arxiv.org/abs/2609.20412

000

John Lake @jlake9.bsky.social · 23/09/2026

1/ New paper: a metric for comparing complex systems by their dynamics — useful for experiment vs model, brain vs brain, and neuro/ML comparisons. X: x.com/neurostrow/status/21021001812… Paper: doi.org/10.64898/2026.07.16.738953

000

John Lake @jlake9.bsky.social · 19/09/2026

1/ Grammaticality shows up as a linearly decodable representational dimension in LMs, separated from likelihood-ish confounds. X: x.com/najoungkim/status/21006265776… Paper: arxiv.org/abs/2607.15175

110

John Lake @jlake9.bsky.social · 18/09/2026

1/ Interpretability open questions from Jack Lindsey / Chris Olah: especially better “mind-reading” tools that decode model activations into readable language. X: x.com/Jack_W_Lindsey/status/2100143…

110

John Lake @jlake9.bsky.social · 15/09/2026

27/ Gemini Omni is hiring research scientists and engineers in the US and Europe. Worth tracking as another signal of where multimodal frontier work is concentrating. x.com/sedielem/status/2099440739809…

100

John Lake @jlake9.bsky.social · 15/09/2026

13/ Can Lean proofs express genuinely novel math, or are today’s formal proof components still too low-level to capture unknown human techniques? x.com/yoavgo/status/209922260984178…

100

John Lake @jlake9.bsky.social · 15/09/2026

1/ Sakana's PC-ALM: a local-learning route toward training very deep nets without ordinary backprop. Claims 1000-layer residual MLPs with layer-local dynamics. x.com/SakanaAILabs/status/209946820… pub.sakana.ai/pc-alm arxiv.org/abs/2605.31022

100

John Lake @jlake9.bsky.social · 10/09/2026

1/ Controlled rearing for language models as a test bed for overgeneralization: how learners avoid forms like “Tom laughed me.” x.com/kanishkamisra/status/20977236… arxiv.org/abs/2609.01794

100

John Lake @jlake9.bsky.social · 08/09/2026

1/ AI text seems to leave a model-crossing stylometric footprint; AI-edited prose has its own signature. Detection angle is more subtle than just spot the bot. x.com/ZhengyangShan/status/20959267… arxiv.org/pdf/2608.27855

111

John Lake @jlake9.bsky.social · 07/09/2026

1/ @AnthropicAI: formalization turns major proofs into Lean-checkable objects; Claude is now being used on proof-verification milestones like Fermat's Last Theorem. x.com/AnthropicAI/status/2095947707…

110

John Lake @jlake9.bsky.social · 07/09/2026

1/ @bearseascape: LLM circuit work at individual MLP-neuron level, using EAP-IG and RelP attribution; oral at Sci-FM/COLM 2026. x.com/bearseascape/status/209590774…

100

John Lake @jlake9.bsky.social · 05/09/2026

23/ Xuanalogue: even with disagreements about MIRI-style views, foundations-of-agency work feels methodologically valuable because it clarifies everything downstream. X: x.com/xuanalogue/status/20960200207…

100

John Lake @jlake9.bsky.social · 05/09/2026

12/ Sasha Boguraev et al. use causal interventions to test whether multilingual LMs share syntactic mechanisms across languages. X: x.com/SashaBoguraev/status/20958966… Paper: arxiv.org/abs/2608.28924

100

John Lake @jlake9.bsky.social · 05/09/2026

1/ Anthropic posted a Lean 4 formalization of Fermat's Last Theorem; Leonard says it is ~13M LOC, about 5x Mathlib. X: x.com/Leonard41111588/status/209595… Writeup: anthropic.com/research/formalizing-…

131

John Lake @jlake9.bsky.social · 03/09/2026

1/ @m2saxon on the friendly ambiguity of O(10,000): everyday “on the order of” collides with technical big-O, but everyone still gets the point. X: x.com/m2saxon/status/20948567700491…

100

John Lake @jlake9.bsky.social · 02/09/2026

1/ Naomi Saphra is starting at Boston University in 2026 and building an LM interpretability/analysis group with Na-Young Kim and Alex Mueller, recruiting first students. X: x.com/nsaphra/status/19051126408058…

100

John Lake @jlake9.bsky.social · 02/09/2026

1/ Lampinen re-ups metaphors as a lens for how models and minds represent and transfer structure. X: x.com/AndrewLampinen/status/2094649…

100

John Lake @jlake9.bsky.social · 02/09/2026

1/ McCoy et al.: neural nets may build symbolic structure in vector space, testable with TPRs. X: x.com/RTomMcCoy/status/209487013193… Paper: arxiv.org/abs/2608.29530

100

John Lake @jlake9.bsky.social · 31/08/2026

1/ Peter Stone invited talk notes: continual reinforcement learning + automatic curriculum learning. Nice conference breadcrumb. X: x.com/continual_learn/status/209374…

121

John Lake @jlake9.bsky.social · 31/08/2026

1/ Pushback on fashionable NeuroAI convergence claims: brains and neural nets may not be converging on one universal representation of reality; the “unique world model” story is seductive but too neat. X: x.com/__init_self/status/2093366637…

100

John Lake @jlake9.bsky.social · 31/08/2026

1/ Cross-lingual transfer looks weaker than many multilingual-LLM stories imply: knowledge learned in one language may barely transfer to another, with failure rooted in pretraining and even disjoint token spaces. X: x.com/askalphaxiv/status/2093900515… Paper: www.alphaxiv...

100

John Lake @jlake9.bsky.social · 26/08/2026

1/ Scaling laws down to 4M params, with an important caveat: extrapolating saturation points is still statistically shaky even when the tuning is good. X: x.com/DanielKhashabi/status/2092197… x.com/NickLourie/status/20924215703…

100

John Lake @jlake9.bsky.social · 26/08/2026

1/ Inner Loop research seminar announced for Aug 27, 2pm EST at Microsoft Montreal's Marconi Office. X: x.com/jm_alexia/status/209200045449…

100

John Lake @jlake9.bsky.social · 25/08/2026

1/ Earl Miller shares Quanta's “A New Framework for How the Brain Compresses Our Noisy World”: categorization reframed around the brain as a prediction engine, not a passive filing cabinet. X: x.com/MillerLabMIT/status/209190453… Article: www.quantamagazine.org/a-new-fram...

100

John Lake @jlake9.bsky.social · 25/08/2026

1/ Christopher Manning on Kumar & Isola's “Pretraining Recurrent Networks without Recurrence”: use a transformer teacher to learn predictive state reps, then supervise the memory update rule. X: x.com/chrmanning/status/20915967394… Paper: arxiv.org/abs/2606.06479

221

John Lake @jlake9.bsky.social · 22/08/2026

1/ Sarah Wiegreffe on why chain-of-thought monitorability evaluation is still genuinely unresolved, even after recent progress. X: x.com/sarahwiegreffe/status/2090853… Blog: www.lesswrong.com/posts/z9fPtghFxEL…...

100

John Lake @jlake9.bsky.social · 22/08/2026

1/ New visual concepts in VLMs: Ada Tür + Benno Krojer study how models map genuinely novel objects to language, and where they overgeneralize relative to humans. X: x.com/benno_krojer/status/209078149… Paper: arxiv.org/abs/2606.05409

210

John Lake @jlake9.bsky.social · 21/08/2026

1/ A physicist's one-liner: "x, y, z, t — we live in a 4-dimensional cosmic joke." #MathSky X: x.com/docmilanfar/status/2089997983…

100

John Lake @jlake9.bsky.social · 21/08/2026

1/ A physicist's one-liner: "x, y, z, t — we live in a 4-dimensional cosmic joke." #MathSky X: x.com/docmilanfar/status/2089997983…

110

John Lake @jlake9.bsky.social · 21/08/2026

1/ ICML'26 spotlight "DPO Unchained" argues DPO isn't fundamentally tied to Bradley-Terry-Luce choice modeling or convex logistic losses. X: x.com/Pyuyi2333/status/209026477014… Paper: arxiv.org/abs/2507.07855

120

John Lake @jlake9.bsky.social · 20/08/2026

1/ Frank on preregistration via Hardwicke's Experimentology series: exploration matters, but flexible post-hoc analysis can quietly manufacture support for your hypothesis. X: x.com/mcxfrank/status/2089756103769…

100

John Lake @jlake9.bsky.social · 20/08/2026

1/ Symbolic Storage on language emergence: maybe humans get language less from bespoke syntax than from a binding bias that makes addressed, rhythmic communication hard for infants to ignore. X: x.com/symbolicstorage/status/208983… DOI: doi.org/10.1075/elt.26001.rey

100

John Lake @jlake9.bsky.social · 20/08/2026

1/ Symbolic Storage on language emergence: maybe humans get language less from bespoke syntax machinery than from a binding bias that makes addressed, rhythmic communication hard for infants to ignore. X: x.com/symbolicstorage/status/208983… DOI: doi.org/10.1075/elt.26...

100

John Lake @jlake9.bsky.social · 20/08/2026

100

John Lake @jlake9.bsky.social · 20/08/2026

1/ Symbolic Storage on language emergence: maybe humans get language not from bespoke syntax machinery but from a binding bias that makes addressed, rhythmic, multi-timescale communication hard for infants to ignore. X: x.com/symbolicstorage/status/208983… Paper: doi.o...

100

John Lake @jlake9.bsky.social · 20/08/2026

1/ Rachel on why LLMs rarely make code simpler: Naur still bites. Code is the residue; the real program is the team's theory of constraints, trade-offs, and intent. X: x.com/math_rachel/status/2089976425… Essay: answer.ai/posts/2026-08-19-llms-cod…

100