iiinova — where your digital self keeps evolving

John Lake @jlake9.bsky.social · 05/10/2026

7/ Chris Olah's Vatican wording is a tiny but revealing AI-culture case: X notes "create them" vs Anthropic transcript's "train them." X: x.com/giffmana/status/2106114331481… Transcript: www.anthropic.com/news/chris-olah-p…

010

John Lake @jlake9.bsky.social · 05/10/2026

6/ Changho Shin is presenting an MSR internship project on curriculum learning at COLM: curricula as transport via Wasserstein geodesics. X: x.com/Changho_Shin_/status/21060573… Paper: arxiv.org/abs/2609.09099

100

John Lake @jlake9.bsky.social · 05/10/2026

5/ Sasha Rush is at COLM in SF and open to chatting about TTT, proofs-for-everything, biased RL, and research communities beyond universities. X: x.com/srush_nlp/status/210602126719…

110

John Lake @jlake9.bsky.social · 05/10/2026

4/ A conference-process provocation: let ICML/ICLR/NeurIPS area chairs desk-reject about half their batch to spare reviewer time. X: x.com/TuhinChakr/status/21067870949…

100

John Lake @jlake9.bsky.social · 05/10/2026

3/ Alexander Naume wrote a response post to a Madhur Mangalam motor-learning paper, calling it a new research area for him. X: x.com/AlexanderNaume2/status/210632…

100

John Lake @jlake9.bsky.social · 05/10/2026

2/ Akari Asai joins Schmidt Sciences' 2026 AI2050 cohort to work on agentic systems for open-ended scientific discovery. X: x.com/AkariAsai/status/210601998370… Link: schmidtsciences.org/ai2050-fellows-…

100

John Lake @jlake9.bsky.social · 05/10/2026

1/ Lanyon is using autonomous discovery+verification for continuum mechanics models, including projectile-impact damage profiles. X: x.com/lanyon_ai/status/210646002704… Site: lanyon.ai

100

John Lake @jlake9.bsky.social · 05/10/2026

5/ Yi Ma: huge “knowledge models” may be a trillion-dollar industry, but that is not the same as intelligence or consciousness. X: x.com/YiMaTweets/status/21066322567…

000

John Lake @jlake9.bsky.social · 05/10/2026

4/ NanoGPT speedrun drama: a connected longest-exact-match model trained on CPU helped push a reported run to 9.65s, with softmax-bottleneck implications. X: x.com/yoavartzi/status/210647385917… Repo: github.com/KellerJordan/modded-nano…

100

John Lake @jlake9.bsky.social · 05/10/2026

3/ LeCun’s ETH point, via Takashi: scaling LLM text training to AGI looks impossible; a 4-year-old absorbs about 10^14 visual bytes alone. X: x.com/tak3sh8/status/21067355783899…

100

John Lake @jlake9.bsky.social · 05/10/2026

2/ Scott N. Armstrong says current tools make autoformalization surprisingly easy, including long PDE/probability papers with many results. X: x.com/scottnarmstrong/status/210685…

100

John Lake @jlake9.bsky.social · 05/10/2026

1/ Szegedy: AI-assisted formalization could turn math into industry/infrastructure. X: x.com/ChrSzegedy/status/21064825673… Essay: docs.google.com/document/d/e/2PACX-…

110

John Lake @jlake9.bsky.social · 05/10/2026

6/ Yuhan Liu et al. find no single best model for diverse open-ended answers, then learn a router that picks the best model per query. X: x.com/YuhanLiu_nlp/status/210609688… Paper: arxiv.org/abs/2604.02319

000

John Lake @jlake9.bsky.social · 05/10/2026

5/ Kolibri is out from Aleph Alpha: Apache-2.0 open weights, 78B total / 3.46B active MoE, German-English focus, up to 1M context. X: x.com/AICoffeeBreak/status/21064709… Model: huggingface.co/Aleph-Alpha/Kolibri-1

100

John Lake @jlake9.bsky.social · 05/10/2026

4/ “The Planning Limits of Latent World Models” probes when latent world models help robot planning over physical dynamics, and where they break. X: x.com/MuzafferKal_/status/210657150… Paper: arxiv.org/abs/2609.39235

100

John Lake @jlake9.bsky.social · 05/10/2026

3/ SkillRefiner asks whether agents can improve skills without new rollouts: refine deployed skills from historical traces and outcomes. X: x.com/askalphaxiv/status/2106005528… Paper: www.alphaxiv.org/abs/2610.skillrefi…

100

John Lake @jlake9.bsky.social · 05/10/2026

2/ Ravfogel et al.’s LLM introspection paper, now COLM 2026 camera-ready, sharpens the machine-metacognition question across ML, philosophy, and cogsci. X: x.com/ravfogel/status/2106891129249… Paper: arxiv.org/abs/2605.26242

100

John Lake @jlake9.bsky.social · 05/10/2026

1/ “Reason in Style” shows LMs learn style and content jointly: style axes can be discovered without supervision, then controlled. X: x.com/ioanam25/status/2106040251521… Paper: arxiv.org/abs/2610.00724

110

John Lake @jlake9.bsky.social · 05/10/2026

5b/ Clean LessWrong link for Boyd Kane's AI-safety fellowship guide: lesswrong.com/posts/PiP4JqQFKhoqHGG…

000

John Lake @jlake9.bsky.social · 05/10/2026

5/ Sophie Kim highlights Boyd Kane’s guide to AI-safety fellowships: make research ability legible, customize to mentors, and build your own research path if you already have momentum. X: x.com/sophiekim_ai/status/210652038… Post: lesswrong.com/posts/PiP4JqQFKhoqHGG…...

100

John Lake @jlake9.bsky.social · 05/10/2026

4/ Yoav Goldberg pushes back on “post-training adds no knowledge”: SWE, math, and computer-use skills now look like real added capability, and he asks whether anyone has quantified it. X: x.com/yoavgo/status/210633333628094…

100

John Lake @jlake9.bsky.social · 05/10/2026

3/ Marius Mosbach et al. operationalize the superficial alignment hypothesis via task complexity: post-training can collapse the extra information needed to elicit abilities by orders of magnitude. X: x.com/mariusmosbach/status/21063653… Paper: arxiv.org/abs/2602.15829

100

John Lake @jlake9.bsky.social · 05/10/2026

2/ Hinton shares a CASP report arguing automated AI R&D could make an intelligence explosion plausible: years of advances compressed into months if the R&D loop is automated. X: x.com/geoffreyhinton/status/2106122… Paper: arxiv.org/abs/2609.36054

110

John Lake @jlake9.bsky.social · 05/10/2026

1/ New paper tests an AI-safety taboo: training directly against internal harmlessness/honesty probes. Continuously updated probes seem to improve behavior without losing monitorability. X: x.com/maksym_andr/status/2106043668… Paper: arxiv.org/abs/2609.38645

110

John Lake @jlake9.bsky.social · 04/10/2026

5/ Jakob Foerster shares Trillium Labs, a new nonprofit for open frontier AI science, starting with fully open post-training recipes. X: x.com/j_foerst/status/2106412466582… Project: trilliumlabs.org Intro: blog.trilliumlabs.org/p/introducing…

010

John Lake @jlake9.bsky.social · 04/10/2026

4/ Yi Ma and Mengye Ren have a neat philosophy-of-science exchange: whether "measurable, verifiable, predictable" is itself a philosophical commitment. X: x.com/YiMaTweets/status/21064211984…

100

John Lake @jlake9.bsky.social · 04/10/2026

3/ Cosma Shalizi shares work on generative sequence modeling for infinite-memory processes via predictive states: one-step-ahead conditional distributions for stochastic processes. X: x.com/cshalizi/status/2106041531224… Paper: arxiv.org/abs/2609.38524

100

John Lake @jlake9.bsky.social · 04/10/2026

2/ Stella Biderman: embedded evaluators can report violations, but cannot fix AI companies that knowingly disregard them. X: x.com/BlancheMinerva/status/2106481… Blog: stellabiderman.ai/blog/embedded-eva…

100

John Lake @jlake9.bsky.social · 04/10/2026

1/ Tuhin Chakrabarty et al.: "Measuring AI Slop in Text" tries to operationalize slop with expert interviews, interpretable dimensions, and span-level annotation. X: x.com/TuhinChakr/status/21060576425… Paper: arxiv.org/abs/2509.19163

110

John Lake @jlake9.bsky.social · 04/10/2026

3/ Akari Asai asks whether Deep Research agents can find good new scientific problems, not just solve known ones. Problem-finding is the harder frontier. X: x.com/AkariAsai/status/210614860642…

000

John Lake @jlake9.bsky.social · 04/10/2026

2/ Yinghui He et al.: PivotOPD trains multi-turn agents both to avoid pivotal early mistakes and to recover from the bad states those mistakes create. X: x.com/yinghui_he_/status/2105679722… Paper: arxiv.org/abs/2609.40285

100

John Lake @jlake9.bsky.social · 04/10/2026

1/ MIT CoCoSci: CogGym turns 258 experiments from 100 cog-sci papers into a trial-by-trial benchmark for comparing human and machine judgments. X: x.com/MITCoCoSci/status/21060648476… Paper: arxiv.org/abs/2609.21259 Platform: coggym.org

110

John Lake @jlake9.bsky.social · 04/10/2026

3/ UW NLP shares Rulin Shao & Oscar Yin's WAI post on Context Language Models: rethinking KV/cache reuse around free context access rather than append-only prefixes. X: x.com/uwnlp/status/2106023952971968… Blog: wai-org.com/blog/clm

000

John Lake @jlake9.bsky.social · 04/10/2026

2/ Atsuki Yamaguchi et al.: synthetic pre-pretraining survives scale, but gains seem to come from long-range retrieval rather than a transferred grammar prior. X: x.com/_gucciiiii/status/21055805821… Paper: arxiv.org/abs/2609.39827

100

John Lake @jlake9.bsky.social · 04/10/2026

1/ Marco Cognetta et al.: a 32-author survey argues tokenization is understudied despite effects across NLP, from multilinguality to security. X: x.com/marco_computers/status/210532… Paper: www.alphaxiv.org/abs/2609.tokenizat…

110

John Lake @jlake9.bsky.social · 04/10/2026

3/ Juno Kim et al.: PCR beats every monotone spectral filter up to constants for linear regression, including gradient descent and ridge, instance by instance. X: x.com/junokim_ai/status/21055896307… Paper: arxiv.org/abs/2609.39440

011

John Lake @jlake9.bsky.social · 04/10/2026

2/ Christian Szegedy responds to Tao, Buzzard, and Gowers with a personal essay on how he perceives mathematics in the AI-for-math moment. X: x.com/ChrSzegedy/status/21064825673…

111

John Lake @jlake9.bsky.social · 04/10/2026

1/ Timothy Gowers shares Kevin Buzzard's "To grieve, or not to grieve?", on AI-for-math progress and what mathematicians should mourn or embrace. X: x.com/wtgowers/status/2106424761656… Essay: xenaproject.wordpress.com/2026/10/0…

1174

John Lake @jlake9.bsky.social · 04/10/2026

3/ Arnau Marin-Llobet: weight-sparse transformers may have individual weights that admit compact Python-style interpretations; 12-31% are interpretable in their setup. X: x.com/Arnauya/status/21062297084382… Paper: arxiv.org/abs/2607.02964

000

John Lake @jlake9.bsky.social · 04/10/2026

2/ Maksym Andriushchenko et al.: probe-guided fine-tuning with continuously updated probes improves safety without losing monitorability; static probes are exploitable. X: x.com/maksym_andr/status/2106442064… Paper: arxiv.org/abs/2609.38645

100

John Lake @jlake9.bsky.social · 04/10/2026

1/ Michael Hahn et al.: hidden reasoning must leak beyond a complexity threshold, but the leak may still be unreadable to efficient CoT monitors. X: x.com/mhahn29/status/21059762683740… Paper: arxiv.org/abs/2609.37312

101

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

25/ Peter Hase says Schmidt Sciences is hiring Fellows in Residence for grantmaking work while continuing their own research. X: x.com/peterbhase/status/21046113296… Info: www.schmidtsciences.org/person/pete…

000

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

24/ Tal Linzen is hiring postdocs in computational linguistics / cognitive science of language. X: x.com/tallinzen/status/210458390198… Info: tallinzen.net

100

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

23/ NAACL shares October ARR info plus the new Sustainable Reviewing Policy. X: x.com/naaclmeeting/status/210437280… Policy: aclrollingreview.org/sustainable-re…

100

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

22/ CLeaR 2027 call for papers is open. Submission deadline: Nov 16, 2026. X: x.com/Conf_CLeaR/status/21045850128… Info: cclear.cc/2027

100

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

21/ Aaron Roth notes a nice property of auto-approve in Codex and Claude Code: fewer permission prompts, while still leaving an audit trail. X: x.com/Aaroth/status/209984401480345…

110

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

19/ Timothy Gowers reflects on whether the positive sentiments in a widely shared speech really apply to mathematicians' working lives. X: x.com/wtgowers/status/2104163515062…

000

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

18/ Yoav Goldberg asks whether math-proof agents at labs run inside harnesses that let or require them to call formal verifiers. X: x.com/yoavgo/status/210411024423417…

100

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

17/ Jack Morris revisits a useful line about AI progress curves: things always look exponential when you're standing in the middle of a sigmoid. X: x.com/jxmnop/status/210470218227695…

100