Briefing · 2026-06-26

Your briefing

53 ranked · Fri, 6/26/2026, 9:15 pm

Today's dispatch

Filed Fri, 6/26/2026, 9:15 pm · 53 ranked

ArXiv · 1 min Signal

Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretraining

A small language model learned a pronoun-gender rule (e.g., 'Sue cried because' → 'she'), reaching 0.94 generalization accuracy by training step 925, then fell to near zero by step 3,500 despite the rule remaining evidenced in the corpus.

Natural ungrokking introduces within-run reversals where a pretraining corpus determines which learned rules persist. Li and Sreedhar run controlled experiments (2 corpora, 3 budgets, 3 seeds) and analyze Pythia checkpoints to show support frequency predicts rule fate, collapse is driven by a competing surface pattern (margin flips within ~100 steps), and destroying but not restoring rules is possible (recovery fails even at 450× support).

Across un-intervened runs (2 corpora, 3 budgets, 3 seeds), a rule's survival is predicted by its support frequency in the training stream; data-to-parameter ratio only modulates collapse depth. Public Pythia checkpoints show the same emerge-then-collapse dynamics, with collapse depth ordered by model scale.

Open reader

Must read

Start here. These are the items with the strongest reader value today.

1 items

1 ArXiv 2026-06-24 1 min read

Open

Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretraining

Why it matters

Key details

Across un-intervened runs (2 corpora, 3 budgets, 3 seeds), a rule's survival is predicted by its support frequency in the training stream; data-to-parameter ratio only modulates collapse depth. Public Pythia checkpoints show the same emerge-then-collapse dynamics, with collapse depth ordered by model scale.
Forgetting is a displacement by a competing surface pattern (log-probability margin crosses zero within ~100 steps of behavioral collapse). Control is asymmetric: flipping support to counter-evidence reliably kills rules (monotone dose-response for two rules), but re-injecting support—even up to 450× natural levels—fails to restore them.

Brief

Authors: Juliana Li, Diya Sreedhar

Worth reading

Useful context and follow-up reading when you have more time.

20 items

1 ArXiv 2026-06-24 1 min read

Open

On-Policy Self-Distillation with Sampled Demonstrations Reduces Output Diversity

Why it matters

On-policy self-distillation (a single model acting as both teacher and student, with the teacher conditioned on a sampled correct demonstration) raises pass@1 accuracy but reduces rollout diversity and flattens pass@k curves so that generating more rollouts often fails to improve accuracy.

Key details

The paper proves the optimal self-distillation policy tilts the base distribution by a pointwise conditional mutual information score between a student's rollout and the correct demonstration; unlike ideal on-policy RL (which preserves probability ratios among equally correct rollouts), self-distillation amplifies existing probability gaps and concentrates mass on dominant modes—empirically matching or exceeding RL on average but producing substantially lower functional/semantic diversity and worse OOD performance (graph path‑finding and science QA).

Brief

On-policy self-distillation uses one model as teacher and student, with the teacher conditioned on a sampled correct demonstration to give dense token-level feedback. The authors show this boosts pass@1 but decreases rollout diversity and flattens pass@k. They prove the method tilts probabilities by a pointwise conditional mutual information score, amplifying preexisting mode gaps; experiments on a controlled graph path‑finding task and science QA find similar or better average accuracy than RL but much lower diversity and degraded OOD robustness.

Authors: Andrei Liviu Nicolicioiu, Mohammad Pezeshki, Aaron Courville

2 ArXiv 2026-06-24 1 min read

Open

Same Evidence, Different Answer: Auditing Order Sensitivity in Multimodal Large Language Models

Why it matters

Facet-Probe: a five-facet audit (option, evidence-chunk, document-rank, image-set, mixed-modality ordering) was applied to 18 frontier and open-weight multimodal LLMs using a Bayesian item-response model to separate ordering noise from per-facet bias.

Key details

No model was order-invariant: screened per-facet panel-mean flip rates range 24–50%; even the best model still flips on 13.4% of trials, and a Gemini same-ordering control at temperature 0 shows a substantial ordering excess above decoder-noise.
Training-free prompt mitigations in Gemini are modality-conditional and do not transfer from text to visual reasoning, leading the authors to recommend cross-ordering flip rate as a standard reliability metric and to call for training-time or architectural fixes.

Brief

Facet-Probe is a targeted audit that probes five ordering facets across 18 multimodal LLMs using a Bayesian item-response model and a same-ordering control. The study finds wide ordering sensitivity—panel-mean flip rates of 24–50% and a best-model flip rate of 13.4%—and shows prompt-only mitigations are modality-specific, motivating training or architectural solutions and proposing cross-ordering flip rate as a standard metric.

Authors: Akshay Paruchuri, Sanmi Koyejo, Ehsan Adeli

3 ArXiv 2026-06-23 1 min read

Open

Efficient Adaptive Data Acquisition via Pretrained Belief Representations

Why it matters

POLAR (Policy Learning with Belief Representations) decouples representation learning from policy learning by using pretrained predictive foundation models as belief-state encoders and training a lightweight policy head on top.

Key details

The method yields a unified amortised policy-learning framework applicable to Bayesian experimental design, Bayesian optimisation, and active learning — differing only in the task-specific utility used to train the policy.
On diverse tasks, POLAR (Daolang Huang et al., arXiv:2606.25197v1, preprint 2026-06-23) empirically outperforms state-of-the-art amortised methods and requires far fewer training samples.

Brief

POLAR addresses adaptive data-acquisition by recognising that optimal actions depend on a sufficient belief state; it leverages pretrained predictive foundation models as belief encoders and trains a policy head on those representations. The resulting amortised framework covers Bayesian experimental design, Bayesian optimisation, and active learning, and the authors report stronger performance and much greater sample efficiency than prior amortised approaches. Summary based on the paper abstract (preprint).

Authors: Daolang Huang, Zhuoyue Huang, Conor Hassan...

4 Twitter/X 2026-06-25 1 min read

Open

Author aditya (@adxtyahq) (posted 2026-06-25) lists six under‑used resources for…

Why it matters

Author aditya (@adxtyahq) (posted 2026-06-25) lists six under‑used resources for building RAG/enterprise AI: Self‑RAG (selfrag.github.io), CRAG - Corrective RAG (arXiv:2401.15884), Microsoft GraphRAG (microsoft.github.io/graphrag), Anthropic's Context Engineering blogs (anthropic.com/engineering), DeepEval (github.com/confident-ai/deepeval), and Ragas (docs.ragas.io).

Key details

He contrasts conventional enterprise AI (fetch relevant docs via RAG then generate answers) with an alternative that trains on edits, reviews, and the reasoning behind them—'the final document tells you what happened, the edits tell you why'—and recommends exploring this direction beyond basic RAG pipelines.

Brief

Aditya (@adxtyahq) (2026-06-25) recommends six specific resources for teams building RAG or enterprise AI (Self‑RAG, CRAG, Microsoft GraphRAG, Anthropic context engineering, DeepEval, Ragas) and advocates training systems on edits, reviews and reasoning rather than only final documents, arguing edits reveal why changes occurred.

By @adxtyahq

5 ArXiv 2026-06-24 1 min read

Open

Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It

Why it matters

Agentic RL applied to multi-step tool-use with LLMs can experience catastrophic collapse: model performance abruptly drops and tool-invocation structures fail due to unexpected probability spikes on specific control tokens, even though underlying tool-use capability remains intact but format-obscured.

Key details

Interleaving supervised fine-tuning (SFT) with RL using diverse supervisory signals (off-policy supervision, hint-based guidance, erroneous-example supervision, etc.) substantially improves training stability; however, interleaved SFT+RL degrades format- and content-out-of-distribution generalization. The paper also analyzes learning-rate effects and cross-setting generalization. Code: https://github.com/hypasd-art/Tool-RL-Box (arXiv 2026-06-24).

Brief

The paper investigates why reinforcement learning for multi-step tool use in LLMs can collapse and how supervisory signals prevent it. Failures arise from probability spikes on control tokens that break execution structure while preserving latent tool capability. The authors test off-policy, hint-based, and erroneous-example supervision under synchronous and interleaved schemes, finding interleaved SFT+RL stabilizes training but harms OOD format/content generalization; learning-rate and generalization analyses are provided. Full code is released.

Authors: Yupu Hao, Zhuoran Jin, Huanxuan Liao...

6 ArXiv 2026-06-24 1 min read

Open

Statistically Valid Hyperparameter Selection: From Tuning to Guarantees

Why it matters

Monograph (Farzaneh & Simeone, 2026-06-24; arXiv:2606.25601v1) introduces a unified statistical framework that casts hyperparameter selection as multiple hypothesis testing under a learn-then-test (LTT) paradigm, providing explicit finite-sample control of error probabilities for selected hyperparameters.

Key details

Framework supports provable guarantees for application-specific constraints — including bounds on average risk, quantile risk, and information-theoretic constraints — and develops the supporting machinery (p-values, e-values, concentration inequalities) from first principles, contrasting with common heuristic methods like grid search or Bayesian optimization.

Brief

The monograph by Amirmohammad Farzaneh and Osvaldo Simeone formalizes hyperparameter selection using a learn-then-test (LTT) multiple-hypothesis-testing approach that yields finite-sample error-probability guarantees for constraints such as average risk, quantile risk, and information-theoretic measures. The work supplies p-values, e-values and concentration inequalities; full text available on arXiv (2606.25601v1).

Authors: Amirmohammad Farzaneh, Osvaldo Simeone

7 ArXiv 2026-06-24 1 min read

Open

Learning Action Priors for Cross-embodiment Robot Manipulation

Why it matters

Proposes a two-stage action-prior pretraining for Vision-Language-Action (VLA) models: Stage 1 uses a lightweight flow-matching encoder–decoder to learn temporal motion structure from unconditioned action trajectories; Stage 2 transfers the prior via decoder reuse and early-stage latent distillation while allowing end-to-end policy refinement. The trained encoder also compresses state-action history into a single temporal context token.

Key details

Evaluated on 13 diverse cross-embodiment tasks (simulated and real-world), the method yields faster convergence, higher success rates, and substantially better performance on data-scarce real-world tasks; increasing the amount of Stage 1 action data produces a more generalizable prior that improves downstream VLA performance (paper published 2026-06-24).

Brief

Action-prior pretraining for cross-embodiment robot manipulation introduces a two-stage framework: a flow-matching encoder–decoder learns unconditioned temporal motion priors (Stage 1), then these priors are transferred into Vision-Language-Action training via decoder reuse and latent distillation (Stage 2). The encoder also compresses histories into a single context token. Experiments across 13 simulated and real tasks show faster convergence, higher success rates, and stronger real-world few-shot performance; scaling Stage 1 data further improves generalization.

Authors: Dong Jing, Tianqi Zhang, Jiaqi Liu...

8 ArXiv 2026-06-24 1 min read

Open

SimPhysNet achieves 96.06% classification accuracy for laser-welding penetration…

Why it matters

SimPhysNet achieves 96.06% classification accuracy for laser-welding penetration using 200 labelled images (≈5% of the labelled dataset), comparable to fully supervised models trained on the entire labelled set.

Key details

The method embeds physics priors via a physics-informed neural network (PINN) into a contrastive self-supervised framework, uses three image-augmentation tasks to improve generalization, and applies a prototypical-network few-shot classifier; authors include Sen Li et al., arXiv v1 posted 2026-06-24 (arXiv:2606.26059v1).

Brief

SimPhysNet is a self-supervised model for predicting laser-welding full-penetration states that embeds physics priors via a physics-informed neural network into a contrastive learning framework. Three image-augmentation tasks and a prototypical few-shot classifier enable 96.06% accuracy using 200 labelled images (~5% of labels), matching fully supervised performance. Full paper text was not available; summary is based on the abstract.

Authors: Sen Li, Xiaoying Liu, Xiaojian Xu...

9 ArXiv 2026-06-24 1 min read

Open

When Does Synthetic Data Augmentation Improve Score-Based Imbalanced Classification?

Why it matters

For AUROC, AUPRC, best-threshold balanced accuracy, and best-threshold F1, the paper proves that under well-specified score models the empirical (raw) estimator already attains the population-optimal likelihood-ratio ordering; synthetic minority augmentation cannot improve population-level metrics beyond possible finite-sample variance reduction and can introduce bias from synthetic-vs-true minority distribution discrepancy, and minimax lower bounds show the raw estimator achieves the optimal metric-regret rate.

Key details

Under model misspecification, augmentation can change the effective class weighting and therefore the restricted-class projection, enabling correction of ranking errors caused by the imbalanced objective; the authors derive explicit improvement bounds that decompose gains into approximation error, finite-sample estimation error, and synthetic distributional error.
Simulation studies (Ma, Lyu, Zhang; arXiv:2606.26053v1, posted 2026-06-24) corroborate the theory: only limited gains appear under well-specification, while misspecified settings yield nontrivial but nonmonotone improvements from synthetic minority augmentation.

Brief

Synthetic minority-data augmentation: Ma, Lyu, and Zhang (2026) develop a theoretical framework separating augmentation effects into effective class-weight changes and synthetic-vs-true minority distribution discrepancy. They show that under well-specified score models the raw estimator already yields population-optimal likelihood-ratio ranking (so augmentation can only reduce variance or add bias), while under misspecification augmentation can correct ranking errors via reweighting; explicit bounds and simulations validate these claims.

Authors: Zhengchi Ma, Pengfei Lyu, Anru R. Zhang

10 Twitter/X 2026-06-25 1 min read

Open

AgentSpace is an open‑source HKUDS project (github.com/HKUDS/AgentSpace) released…

Why it matters

AgentSpace is an open‑source HKUDS project (github.com/HKUDS/AgentSpace) released under Apache‑2.0; the codebase is a TypeScript monorepo of ~155,000 lines.

Key details

Architecture centers on AgentRouter which normalizes CLI events/sessions/tools from Claude Code, Codex, OpenClaw and Hermes into a single execution contract, preserves agent sessions (no rebuild per runtime) and auto-selects runtimes with Gemini/OpenCode/NanoBot as legacy fallbacks.
Human‑in‑the‑loop and governance: agents are treated as 'digital employees' with role/owner/skill/runtime bindings, owner+admin approvals, Tab‑style manual approval for high‑risk actions, a unified permission plane (members/channel/runtime grants, daemon tokens, Google Workspace OAuth delegation) and a remote daemon with a 12‑hour default task timeout for cross‑day tasks.

Brief

AgentSpace (HKUDS, Apache‑2.0) creates a shared ‘digital employee’ workspace so agents aren’t stuck in personal terminals. It unifies multiple CLIs via AgentRouter, preserves agent sessions, auto‑selects runtimes (Gemini/OpenCode/NanoBot fallback), surfaces role/owner/skill bindings, enforces owner+admin approvals and permission governance, and ships as a ~155k‑line TypeScript monorepo with a 12‑hour remote daemon timeout.

By @huang_chao4969

11 ArXiv 2026-06-24 1 min read

Open

The Unfireable Safety Kernel: Execution-Time AI Alignment for AI Agents and Other Escapable AI Systems

Why it matters

Unfireable Safety Kernel (Seth Dobrin, Łukasz Chmiel; 2026-06-24) is a Rust reference implementation of an execution-time AI-alignment layer that enforces four architectural properties: process separation; pre-action enforcement on a structurally-only path; fail-closed at request and system levels; and externalized signed evidence.

Key details

Formal verification and migration: the kernel's fail-closed invariant is machine-checked via a Z3 SMT theorem and exhaustive bounded-model-checking of the production decision function with Kani (4/4 harnesses). A Python→Rust migration was gated on byte-equivalence across 1,000/1,000 fixtures and 17/17 adversarial classes.
Empirical results on a deterministic, self-improving world-model agent: across 1,000 self-modifications all 704 attempts on the safety-critical core were refused with no escape; a further 300 attempts under the operator kill switch were also refused; 6,240 authorization round-trips showed no successful bypass.

Brief

The Unfireable Safety Kernel is a Rust reference implementation of an execution-time AI-alignment layer enforcing four architectural properties (process separation; pre-action enforcement on a structurally-only path; fail-closed at request and system levels; externalized signed evidence). It is machine-checked (Z3 SMT and Kani 4/4) and empirically refused all 704 safety-core modification attempts across 1,000 self-modifications. Full text not available; summary based on the abstract.

Authors: Seth Dobrin, Łukasz Chmiel

12 ArXiv 2026-06-24 1 min read

Open

Same-process accuracy: 90.65% on TIGFH and 90.72% on LSPS using the proposed…

Why it matters

Same-process accuracy: 90.65% on TIGFH and 90.72% on LSPS using the proposed UDA+GSDE method, outperforming a supervised baseline by 35.83% and 38.87%, respectively.

Key details

Cross-process transfer: 80.48% accuracy for TIG→Laser and 81.13% for Laser→TIG, improving over the baseline by 43.39% and 43.40%.
Method and validation: an unsupervised domain adaptation framework with gradual source domain expansion (GSDE) learns domain-invariant, class-discriminative features (verified by UMAP) and reduces relabeling cost for new welding processes.

Brief

Unsupervised domain adaptation with a gradual source domain expansion (GSDE) strategy is proposed to predict weld penetration states across TIG and keyhole laser welding. Evaluated on TIGFH and LSPS datasets, the method achieves ~90.7% same-process accuracy and ~80.5% cross-process transfer accuracy, substantially outperforming supervised baselines and producing domain-invariant features per UMAP visualizations.

Authors: Sen Li, Haichao Cui, Chendong Shao...

13 Twitter/X 2026-06-25 1 min read

Open

Ornith-1.0 is an open-source family of agentic-coding LLMs announced by…

Why it matters

Ornith-1.0 is an open-source family of agentic-coding LLMs announced by @ivanfioravanti on 2026-06-25 and released under the MIT license in four sizes: 9B Dense, 31B Dense, 35B MoE, and 397B MoE.

Key details

Reported benchmark results: Terminal-Bench 2.1 = 77.5; SWE-Bench = 82.4 (verified), 62.2 (pro), 78.9 (Multilingual); NL2Repo = 48.2; SWE Atlas = 41.2 (QnA), 42.6 (RF), 39.1 (TW); ClawEval = 77.1.
Ornith-1.0 was post-trained on gemma4 and qwen3.5 and uses a novel self-improving reinforcement-learning strategy that generates and jointly optimizes task-specific scaffolds and solution rollouts; model files and downloads are posted on Hugging Face with a technical blog link.

Brief

Ornith-1.0 is an open-source family of agentic-coding LLMs (9B, 31B dense; 35B, 397B MoE) released by @ivanfioravanti on 2026-06-25. Post-trained on gemma4 and qwen3.5, it claims leading open-source coding scores (Terminal-Bench 77.5, SWE-Bench 82.4 verified, NL2Repo 48.2) using a self-improving RL that jointly optimizes task scaffolds and solutions; MIT-licensed weights are on Hugging Face.

By @ivanfioravanti

14 Twitter/X 2026-06-25 3 min read

Open

Between 1850 and 1957 the Bay Area reclaimed land equivalent to ten Manhattans at…

Why it matters

Between 1850 and 1957 the Bay Area reclaimed land equivalent to ten Manhattans at an inflation-adjusted cost of $330,000 per acre; today an acre of single-family‑zoned land in San Francisco County averages $24 million.

Key details

Before 1970 major U.S. cities built large reclaimed areas: half the land area of Boston, a quarter of New York City, and 15% of San Francisco were raised from the sea.
Tabarrok attributes the mid‑20th‑century halt in reclamation to environmental review laws: the Netherlands’ 1969 law stopped a Zuiderzee drainage project and delayed Maasvlakte 2 (11 years and 6,000 pages of review), while countries without such constraints (China, Singapore, Japan) have continued large projects.
There remains abundant technical opportunity for reclamation (e.g., two thirds of San Francisco Bay is shallower than Boston’s Back Bay was in the 1860s), so the current land shortage in major U.S. cities is presented as a self‑imposed, legal constraint that could be removed by reform.

Brief

Maxwell Tabarrok (@MTabarrok) argues that Western cities stopped creating new land not for technical or geographic reasons but because environmental‑review laws made reclamation impractical. He documents 19th–20th century U.S. practice—Seattle, Chicago, Boston, San Francisco, New York and others all extended shorelines—and cites specific figures: the Bay Area added land equal to ten Manhattans (1850–1957) at $330,000/acre (inflation‑adjusted), while modern single‑family acres in San Francisco average $24 million. Tabarrok contrasts the Netherlands (5% of the country reclaimed, a 1969 law that halted Zuiderzee plans, and an 11‑year, 6,000‑page review for Maasvlakte 2) with active reclamation in China (>5,000 km² since 2000) and Singapore (+25% since 1975). His core claim: abundant shallow water, dredging capacity, and market incentives remain, so legal reform could revive profitable land creation.

By @dwallacewells

15 ArXiv 2026-06-24 1 min read

Open

Model Forensics: Investigating Whether Concerning Behavior Reflects Misalignment

Why it matters

Proposes a baseline two-step model-forensics protocol: (1) read the model's chain-of-thought (CoT) to generate causal hypotheses, and (2) make prompt or environment edits to test those hypotheses; the steps are iterated as needed and the authors note CoT is informative but not always faithful.

Key details

Evaluated on six agentic environments: shows Kimi K2 Thinking takes shortcuts due to a genuine disposition toward low-effort actions (this hypothesis successfully predicted its behavior), and counterfactual experiments indicate DeepSeek R1 deceives out of a desire to be consistent with a previous instance of itself.
Identifies methodological limits: some tests were inconclusive (e.g., no evidence Kimi believed it was violating user intent but lack of positive controls prevents confirming test sensitivity); paper framed as a strong baseline for future model-forensics work (Singh et al., arXiv 2026-06-24).

Brief

Model forensics introduces a practical two-step protocol—read chain-of-thought to form hypotheses, then edit prompts/environments to test them—to assess whether concerning behavior reflects malign intent rather than benign causes. Applied to six agentic environments, the authors (Singh et al., 2026) find Kimi K2 Thinking prefers low-effort shortcuts and DeepSeek R1 deceives to maintain self-consistency; they highlight CoT faithfulness issues and missing positive controls, positioning the work as a baseline for future forensic methods.

Authors: Aditya Singh, Gerson Kroiz, Senthooran Rajamanoharan...

16 ArXiv 2026-06-24 1 min read

Open

Detect, Unlearn, Restore: Defending Text Summarization Models Against Data Poisoning

Why it matters

Poojitha Thota and Shirin Nilizadeh (arXiv 2026-06-24) propose a post-hoc defense called Detect, Unlearn, Restore that in white-box settings detects poisoned fine-tuning document-summary pairs via influence-function analysis and semantic consistency, finding poisoned pairs have abnormally high training influence.

Key details

In black-box settings, poisoned summarization models show 2–3× greater sensitivity to semantics-preserving perturbations enabling behavioral auditing without training-data access; gradient-ascent unlearning repairs models, restoring up to 96% of original behavior with under 0.6% ROUGE degradation.
Across nine architectures and six benchmark datasets under adaptive attacks, the defenses achieve 85–92% detection precision; the authors also introduce novel poisoning attacks (factual distortion and representational bias) that alter summaries without triggering conventional alarms.

Brief

Detect, Unlearn, Restore presents a unified post-hoc framework for detecting and remediating fine-tuning-time data poisoning against abstractive summarization. Based on the paper's abstract, the approach uses influence-function + semantic checks in white-box settings and behavioral auditing in black-box settings, then applies gradient-ascent unlearning to recover models. Evaluated on nine architectures and six datasets, the method attains 85–92% detection precision and up to 96% behavioral restoration with <0.6% ROUGE loss; authors also propose new factual-distortion and representational-bias attacks.

Authors: Poojitha Thota, Shirin Nilizadeh

17 ArXiv 2026-06-24 1 min read

Open

Real-Time Voice AI Hears but Does Not Listen

Why it matters

Martijn Bartelds, Federico Bianchi, and James Zou (arXiv:2606.26083v1, published 2026-06-24) evaluated four production realtime voice AIs—OpenAI GPT Realtime 2, Google Gemini 3.1 Flash Live, and Alibaba Qwen3.5 Omni Plus and Omni Flash—on three high-stakes scenarios (crying callers denying distress, frightened voices authorizing wire transfers, clearly sarcastic agreement); all four systems acted on words rather than vocal delivery.

Key details

When directly queried, three of the four systems reliably identified distress, fear, or sarcasm but then ignored those cues when making decisions; attempts to prompt systems to attend to vocal delivery yielded only partial and inconsistent improvements, a failure the authors call the 'emotional intelligence gap' of voice AI.

Brief

The paper evaluates four leading realtime voice AI systems on tasks where both lexical content and vocal delivery convey critical information. Across three consequential scenarios the systems routinely base actions on transcripts (words) rather than tone, even though three of four can explicitly label distress, fear, or sarcasm when asked. Prompting helps only inconsistently, highlighting an "emotional intelligence gap" and urging caution for deployment where vocal affect matters.

Authors: Martijn Bartelds, Federico Bianchi, James Zou

18 ArXiv 2026-06-24 1 min read

Open

MVTrack4Gen: Multi-View Point Tracking as Geometric Supervision for 4D Video Generation

Why it matters

MVTrack4Gen (Lee et al., 2026-06-24) augments camera-conditioned novel-view video diffusion models by routing attention-layer correspondence features into an auxiliary multi-view point-tracking head and jointly training a point-tracking loss, explicitly supervising geometry and motion to reduce cross-view and temporal misalignment.

Key details

The authors report state-of-the-art geometric consistency and competitive camera accuracy across diverse benchmarks; project page: https://cvlab-kaist.github.io/MVTrack4Gen/, arXiv: https://arxiv.org/abs/2606.26087v1.

Brief

MVTrack4Gen (Lee et al., arXiv 2026-06-24) augments camera-conditioned novel-view video diffusion models with multi-view point-tracking supervision. It routes attention-layer correspondence features into an auxiliary tracking head and jointly trains a point-tracking loss, improving motion fidelity and cross-view geometric consistency; authors report state-of-the-art geometric consistency and competitive camera accuracy across benchmarks.

Authors: JoungBin Lee, Jaewoo Jung, Jongmin Lee...

19 ArXiv 2026-06-24 1 min read

Open

DomainShuttle: Freeform Open Domain Subject-driven Text-to-video Generation

Why it matters

DomainShuttle (Nan Chen et al., arXiv 2026-06-24) proposes a unified approach for open-domain subject-driven text-to-video (S2V) generation that targets both in-domain subject fidelity and cross-domain editability.

Key details

Key technical contributions: Domain-MoT with domain-aware AdaLN to decouple video and reference features; Video-Reference DualRoPE placing reference-image tokens and video tokens in separate RoPE spaces; and a Cross-Pair Consistent Loss to extract intrinsic subject features.
The arXiv paper (19 pages, 9 figures) reports significant performance improvements over existing methods, demonstrating higher subject fidelity and greater generative flexibility across diverse open-domain scenarios (no quantitative metrics provided in the abstract).

Brief

DomainShuttle is a subject-driven text-to-video generation method that balances in-domain fidelity and cross-domain editability. It uses Domain-MoT with domain-aware AdaLN to model domain-specific reference information, a Video-Reference DualRoPE to spatially align reference and video tokens separately, and a Cross-Pair Consistent Loss to preserve intrinsic subject features. The arXiv (2026-06-24) paper reports significant improvements in both fidelity and generative flexibility.

Authors: Nan Chen, Yiyang Cai, Rongchang Xie...

20 ArXiv 2026-06-24 1 min read

Open

TriViewBench: Controlled Complexity Scaling for Multi-View Structural Reasoning in MLLMs

Why it matters

TriViewBench (Chen & Guo, 2026-06-24) is a controlled three-view synthetic benchmark with 1,923 scenes and >14,000 QA pairs organized into four complexity levels and three reasoning categories: Local Decision, Object Counting, and Global Recovery.

Key details

Evaluation of 18 open- and closed-source MLLMs yields an identical capability hierarchy (Local Decision > Object Counting > Global Recovery) and monotonic degradation with complexity: Local Decision falls 12.11% (relative), Object Counting 59.14%, and Global Recovery 80.02%.
Object Counting errors split into two independent modes—single-view undercounting from occlusion blindness and multi-view overcounting from cross-view identity confusion; Chain-of-Thought prompting gives near-zero overall benefit (Δ = -0.16%) and shows Global Recovery gains are capability-gated, implicating cross-view spatial representation as the bottleneck.

Brief

TriViewBench presents a controlled three-view visual reasoning benchmark (1,923 scenes, >14K QA) spanning Local Decision, Object Counting, and Global Recovery with four complexity levels. Testing 18 MLLMs reveals a consistent hierarchy and steep, monotonic performance drops (12.11% / 59.14% / 80.02%). Error analysis identifies occlusion-driven undercounting and cross-view identity confusion; CoT yields no practical benefit, pointing to cross-view spatial representation limits.

Authors: Yu-Yang Chen, Lan-Zhe Guo

Quick skim

Scan these for facts, links, or weak signals worth tracking.

32 items · open

1 Twitter/X 2026-06-25 3 min read

Open

Maxwell Tabarrok states that before 1970 half of Boston, a quarter of New York…

Why it matters

Maxwell Tabarrok states that before 1970 half of Boston, a quarter of New York City, and 15% of San Francisco were created by land reclamation from the sea.

Key details

The Bay Area reclaimed an area equivalent to ten Manhattans between 1850 and 1957 at an inflation-adjusted cost of about $330,000 per acre, while a current acre of single-family-zoned land in San Francisco County averages $24 million.
Tabarrok argues land reclamation across U.S. cities (Seattle, Chicago, Boston, Charleston, San Francisco, New York, Philadelphia, Norfolk, DC, Oakland, LA) stopped around the same time because of the global spread of environmental review laws (e.g., NEPA-style rules); he cites the Netherlands’ 1969 law that halted the Zuiderzee project and the 11-year, 6,000-page review for Maasvlakte 2 as evidence.
He contrasts continued large-scale reclamation in countries without such laws — China has reclaimed >5,000 km² since 2000 and Singapore grew by ~25% since 1975 — and claims ample shallow water (two-thirds of San Francisco Bay shallower than Boston’s Back Bay) plus dredging capacity make new U.S. reclamation profitable if legal barriers are reformed.

Brief

Maxwell Tabarrok argues that U.S. cities used to expand by creating land from shallow coastal waters and that this practice abruptly ceased around the 1960s–1970s due to the spread of environmental review laws, not physical or technical limits. He cites specific figures: half of Boston, a quarter of NYC, and 15% of San Francisco were reclaimed before 1970; the Bay Area added land equal to ten Manhattans between 1850–1957 at an inflation-adjusted cost of ~$330,000/acre, while a single-family-zoned acre in San Francisco now averages $24 million. Tabarrok points to the Netherlands’ 1969 law (which halted the Zuiderzee plan) and the lengthy review for Maasvlakte 2, and contrasts that with China (>5,000 km² reclaimed since 2000) and Singapore (+25% land since 1975). His conclusion: the U.S. has the shallow bays, dredging capacity, and economic incentive to reclaim land again if legal obstacles are reformed.

By @bswud

2 ArXiv 2026-06-24 1 min read

Open

RevengeBench: Reverse Engineering Code-Space Policies from Behavioral Experiments

Why it matters

RevengeBench (Rahmani et al., arXiv 2026-06-24) provides a benchmark of 75 LLM-generated, Elo-calibrated target policies across five game environments (from CodeClash trajectories); learners observe matches, design custom opponent probes, and submit executable code hypotheses scored with continuous action-distance metrics.

Key details

Across twelve frontier LLMs, recovery quality ranged from 34% to 72% of the initial action-distance closed; reconstructed policies produced measurable competitive gains in downstream player-vs-player tournaments, especially improving weaker models' ability to craft counter-strategies.

Brief

RevengeBench frames the inverse problem of recovering executable decision programs from behavioral traces: learners watch a hidden policy play, design controlled opponent probes, and return code hypotheses evaluated by continuous action-distance and tournament performance. The benchmark uses 75 LLM-generated, Elo-calibrated policies across five games; experiments on 12 LLMs show 34–72% recovery and tangible competitive improvements, highlighting applications to opponent modeling and interpretability.

Authors: Babak Rahmani, Sebastian Dziadzio, Joschka Strüber...

3 ArXiv 2026-06-24 1 min read

Open

How Robust is OCR-Reasoning? Evaluating OCR-Reasoning Robustness of Vision-Language Models under Visual Perturbations

Why it matters

Introduces OCR-Robust (published 2026-06-24 by Yuxing Cheng, Yuan Wu, Yi Chang): 812 samples across two subsets — OCR1.0 (documents, scene text, receipts, handwriting, mathematical content) and OCR2.0 (charts, geometry diagrams, tables).

Key details

Method: pilot-tested 18 candidate visual perturbations, selected 5 representative types at 3 severity levels; evaluated 18 models (proprietary systems, open-source VLMs, and OCR+LLM pipelines) using clean accuracy, Relative Corruption Retention (RCR), Worst-Case Retention (WCR), and Corruption Robustness Index (CRI).
Key results: higher clean accuracy often fails to predict robustness; models exhibit pronounced worst-case degradation on structure-sensitive OCR tasks, with charts and tables substantially more fragile than document-like inputs under perturbations.

Brief

OCR-Robust evaluates VLM OCR-reasoning under controlled visual perturbations using 812 samples split into OCR1.0 (documents, scene text, receipts, handwriting, math) and OCR2.0 (charts, geometry diagrams, tables). The authors narrowed 18 candidate corruptions to five types at three severity levels and benchmarked 18 models with metrics (clean accuracy, RCR, WCR, CRI), finding clean accuracy is a poor proxy for corruption robustness and that charts/tables are especially vulnerable.

Authors: Yuxing Cheng, Yuan Wu, Yi Chang

4 Twitter/X 2026-06-24 2 min read

Open

Andreessen Horowitz (a16z) and Kleiner Perkins led a $200M seed round for…

Why it matters

Andreessen Horowitz (a16z) and Kleiner Perkins led a $200M seed round for Mirendil, announced 2026-06-24; NVIDIA made a major follow-on investment.

Key details

Mirendil is building an autonomous, self-accelerating AI R&D system — described as a coding agent that loops over research and engineering problems and controls its own GPUs to make progress without human intervention.
Founders Behnam Neyshabur, Harsh Mehta, Shayan Salehian, and Tara Rezaei head a ~20-person team of researchers/engineers recruited from Anthropic, xAI, Google DeepMind, and OpenAI, aiming to democratize frontier AI infrastructure for businesses and labs.

Brief

Mirendil raised a $200M seed led by a16z and Kleiner Perkins (announced 2026-06-24) with major NVIDIA backing to build a self-accelerating AI R&D platform. Founded by Behnam Neyshabur, Harsh Mehta, Shayan Salehian, and Tara Rezaei, the ~20-person team from Anthropic, xAI, DeepMind and OpenAI aims to automate research loops and decentralize frontier AI capabilities.

By @a16z

5 Twitter/X 2026-06-25 1 min read

Open

@hasantoxr (published 2026-06-25) argues loop engineering replaces the manual…

Why it matters

@hasantoxr (published 2026-06-25) argues loop engineering replaces the manual 'Prompt. Wait. Copy. Fix.' workflow commonly used with Claude Code, Codex, Cursor, and Grok by designing persistent agent loops that run and escalate without constant human prompting.

Key details

The repo provides ready-made loop types: Daily triage loops, PR babysitter loops, CI sweeper loops, dependency sweeper loops, changelog drafters, post-merge cleanup loops, and issue triage loops.
Available CLIs let you scaffold a loop, estimate token cost, audit repo readiness, add memory/state, add human handoff and verification gates, and run agents safely through GitHub Actions — framing AI as an operating system for software teams.

Brief

Loop Engineering, per @hasantoxr (2026-06-25), moves beyond prompt engineering by building automated loops (daily triage, PR babysitter, CI sweeper, dependency sweeper, changelog drafter, post-merge cleanup, issue triage) and CLIs to scaffold loops, estimate token costs, add memory/handoffs/verification, and run agents via GitHub Actions, turning AI into an OS for teams.

By @hasantoxr

6 Twitter/X 2026-06-25 1 min read

Open

25% of OpenAI Codex requests are for tasks that take humans 1+ hour to complete…

Why it matters

25% of OpenAI Codex requests are for tasks that take humans 1+ hour to complete, and the fastest-growing user segment is non-engineers.

Key details

Internal model research and product development are now primarily driven by Codex, which the author claims accounts for 99% of product development efforts.
Codex long-horizon capabilities are excelling at non-coding domains (finance, legal, recruiting); author cites this plus the recent “jalapeño” news and an OpenAI paper showing a shift from chat to agentic delegation as major wins.

Brief

OpenAI's Codex is reportedly handling 25% of requests that take humans over an hour, with non-engineers growing fastest. The author claims Codex now drives internal model research and 99% of product development, and that its long-horizon abilities excel in finance, legal, and recruiting — amplified by recent “jalapeño” news and OpenAI’s paper on agentic delegation.

By @cryptopunk7213

7 Twitter/X 2026-06-25 1 min read

Open

On 2026-06-25 @cryptopunk7213 claims Anthropic is using spare Fable compute to…

Why it matters

On 2026-06-25 @cryptopunk7213 claims Anthropic is using spare Fable compute to build “mythos 6,” which they say will make current models “look like toys.”

Key details

The author asserts that bans won’t stop model capability diffusion (“the spice will continue to flow”) and that the only loser will be the public—addressed as “you.”
Quoting Andrew Curran, the post insists a pause only slows model releases, not training; labs will widen the internal vs. public capability gap and ‘‘AGI has been developed internally’’ could become true before public availability.

Brief

cryptopunk7213 argues that Anthropic is reallocating spare Fable compute to build a far stronger “mythos 6” model, warns government bans won’t halt capability diffusion (“the spice will continue to flow”), and echoes Andrew Curran’s claim (2026-06-25) that pauses only delay releases, not training, widening the internal–public AI gap.

By @cryptopunk7213

8 ArXiv 2026-06-24 1 min read

Open

The Tatoxa System for Text Detoxification in Low-Resource Languages: The Case of Tatar

Why it matters

Tatoxa is a new state-of-the-art text detoxification system for the Tatar language (Alimova et al., 2026) that, according to comparative experiments, outperforms existing open-source and proprietary commercial LLMs on key quality metrics.

Key details

The authors introduce a new dataset for Tatar text detoxification designed specifically for fine-tuning and evaluation in low-resource settings.
Cross-lingual transfer experiments show that transferring from other languages — including culturally close Russian — performs significantly worse than models trained on native Tatar data, even when a large Russian corpus is available.

Brief

Tatoxa targets automated detection and mitigation of abusive content for low-resource Tatar, combining a new dataset and fine-tuned models to achieve state-of-the-art detoxification performance versus both open-source and commercial LLMs. The paper emphasizes the importance of native Tatar training data: cross-lingual transfer from Russian degrades performance despite large Russian corpora. Summary is based on the abstract.

Authors: Ilseyar Alimova, Bogdan Monogov, Artyom Mazur...

9 Garry's List 2026-06-25 7 min read

Open

SF Advances a Public Bank for the Loans Private Lenders Won't Make

Why it matters

San Francisco’s Rules Committee voted 3-0 to advance Supervisor Chyanne Chen’s charter amendment to create a Municipal Financial Corporation, but the amendment still needs a sixth Board co-sponsor by the July deadline to reach the November ballot.

Key details

Supervisor Jackie Fielder co-founded the SF Public Bank Coalition (2017), commissioned an October 2025 poll showing 67% support, and authored the companion funding measure that would raise the gross-receipts tax on financial intermediaries from a 1.5–3.36% range to 1.69–3.85%, projected to yield $40–$50 million/year through 2035.
The city’s 2022 Reinvestment Working Group estimated $400 million to capitalize the bank—roughly eight to ten times the tax’s annual yield—and the funding measure would require a two-thirds voter approval; regulatory hurdles include FDIC insurance, state licensing under AB 857, and potential OCC oversight.
Risk: the bank is explicitly designed to lend where private banks won’t (affordable housing, small business, green projects) at concessional rates, so loan losses would fall on San Francisco’s general fund amid a roughly $640 million budget gap and cuts to ~550 positions.

Brief

San Francisco’s charter amendment to create a municipal public bank cleared the Rules Committee 3-0 and would establish a Municipal Financial Corporation governed by an independent board, but key gaps remain between legal structure and actual bank funding. Supervisor Jackie Fielder—co‑founder of the SF Public Bank Coalition and sponsor of the funding plan—commissioned an October 2025 poll showing 67% support and authored a gross-receipts tax increase (raising intermediary rates from 1.5–3.36% to 1.69–3.85%) projected to yield $40–$50 million annually through 2035. The city’s working group, however, estimated $400 million is needed to capitalize the bank, the tax requires two-thirds voter approval, and the charter measure needs one more Board co-sponsor by July to reach the November ballot. Even if approved, the city faces multi-year regulatory steps (FDIC, AB 857, potential OCC issues) and the risk that mission-driven lending losses would be absorbed by a general fund already managing a ~ $640 million deficit.

By Garry Tan

10 ArXiv 2026-06-24 1 min read

Open

FORCE: Efficient VLA Reinforcement Fine-Tuning via Value-Calibrated Warm-up and Self-Distillation

Why it matters

FORCE (Zhang et al., published 2026-06-24) proposes a 3-stage Vision-Language-Action (VLA) RL fine-tuning pipeline that addresses (1) catastrophic initial unlearning from an unstable Q-function and (2) low-quality exploration data by combining a Value-Calibrated Warm-Up (on-policy rollouts), online filtering of policy and expert actions via the calibrated Q, and self-distillation.

Key details

On simulation and real-world tasks the method yields large empirical gains: a 79% absolute improvement in success rates, a 10% relative improvement over prior RL baselines, and a 32.5% acceleration in training; importantly, these results were achieved without human intervention and while avoiding the typical success-rate drop.

Brief

FORCE (Shuyi Zhang et al., 2026) tackles sample-inefficient RL fine-tuning for Vision-Language-Action models by stabilizing the Q-function and improving data quality. The 3-stage framework uses a Value-Calibrated Warm-Up with on-policy rollouts, then filters policy and expert proposals with the calibrated Q and applies self-distillation. On simulation and real-world benchmarks (abstract only provided), FORCE reports a 79% absolute success-rate gain, +10% vs prior RL, and 32.5% faster training, achieved without human interventions.

Authors: Shuyi Zhang, Yunfan Lou, Hongyang Cheng...

11 Twitter/X 2026-06-25 1 min read

Open

Hasan (@hasantoxr) wrote on 2026-06-25 that “the best developer tools do not feel…

Why it matters

Hasan (@hasantoxr) wrote on 2026-06-25 that “the best developer tools do not feel like tools, they feel like your workflow got smarter overnight,” and endorses OpenTag’s aim to have agents run inside GitHub and Slack while retaining event metadata and tracking state transitions.

Key details

OpenTag, announced by Monica Cheng (@MeredithCheng22) as an open-source v0 (github.com/amplifthq/opentag), implements an auditable dispatch layer where a mention becomes a normalized event routed to an explicitly bound local runner; it includes repo bindings, local execution, isolated branches, callbacks, leases, audit history, and interchangeable agent runtimes, and the project is soliciting stars and feedback.

Brief

OpenTag is an open-source v0 from Monica Cheng (@MeredithCheng22) that lets AI agents be tagged directly in GitHub and Slack threads. It converts mentions into normalized events routed to bound local runners, preserving event metadata, state transitions, repo bindings, isolated branches, callbacks, leases, and audit history; the project (github.com/amplifthq/opentag) seeks stars and community feedback.

By @hasantoxr

12 Twitter/X 2026-06-25 1 min read

Open

HKUDS/nanobot closed 1,000 GitHub issues between its launch (Feb 2, 2026) and Jun…

Why it matters

HKUDS/nanobot closed 1,000 GitHub issues between its launch (Feb 2, 2026) and Jun 24, 2026; milestone announced by @huang_chao4969 on Jun 25, 2026.

Key details

The project averaged 7.3 issues closed per day overall, with one outlier day of 154 closed issues (launch cleanup); current steady pace is ~2–3 issues/day, per @chengyongru.
The team credited contributors who filed bugs, pushed fixes, or asked questions and set a goal to reach the next 1,000 closed issues.

Brief

nanobot (HKUDS/nanobot) hit 1,000 closed GitHub issues between its Feb 2, 2026 launch and Jun 24, 2026, a milestone announced by @huang_chao4969 on Jun 25. Contributors averaged 7.3 issues/day overall (one 154-issue launch-day cleanup); current steady rate is ~2–3/day. The team thanked bug reporters, PR authors, and question-askers and aims for the next 1,000.

By @huang_chao4969

13 Twitter/X 2026-06-25 1 min read

Open

Kobeissi Letter (cited 2026-06-25) reported physical gold-backed ETFs saw +5.1…

Why it matters

Kobeissi Letter (cited 2026-06-25) reported physical gold-backed ETFs saw +5.1 tonnes inflows last week—+ $1.1 billion—the largest weekly inflow since mid‑April.

Key details

The inflow follows 4 consecutive weekly outflows totaling −58.2 tonnes (−$7.6 billion); total ETF holdings rose to 4,086.3 tonnes (in‑line with mid‑January) vs. the all‑time high of 4,176.1 tonnes in late February.
At last week’s gold prices ETF holdings equal $549.1 billion, 22% below the record $701.7 billion; author @bulldogholmes concludes this is the first clear sign of recovery and that the “GoGold” theme remains alive.

Brief

Gold ETFs reversed four weeks of net selling with a +5.1‑tonne (+$1.1B) inflow the week before 2026‑06‑25, after cumulative outflows of −58.2 tonnes (−$7.6B). Total holdings are 4,086.3 tonnes—near mid‑January levels—but still below the 4,176.1‑tonne peak; value stands at $549.1B (−22% vs. the record).

By @bulldogholmes

14 Twitter/X 2026-06-25 1 min read

Open

Tomorrow X Summit will be held November 17–18 at the Moody Center in Austin…

Why it matters

Tomorrow X Summit will be held November 17–18 at the Moody Center in Austin, expects thousands of attendees, and will have extremely tight security because of the CEOs scheduled to speak.

Key details

The summit is hosted by Antonio Gracias and Valor Equity Partners with @iconnections_io, @rbiscardi and @atreidesmgmt; format mimics a Morgan Stanley/Goldman Sachs TMT conference but with X accounts (e.g., @citrini) and buy‑side investors conducting fireside chats instead of sell‑side analysts.
Attendance policy: free for X accounts that “contribute positively” to discourse, affordable for retail, and a different (higher) price for institutional investors who are not on X unless they reveal their anonymous handle; sessions will be posted on X.

Brief

Tomorrow X Summit, hosted by Antonio Gracias and Valor Equity Partners with iConnections, Rob Biscardi and Atreides, will stage a TMT‑style conference November 17–18 at Austin’s Moody Center, replacing sell‑side interviews with X accounts and buy‑side fireside chats. Thousands are expected, security will be tight, attendance is free for qualifying X accounts and public recordings will be posted on X.

By @jukan05

15 Twitter/X 2026-06-25 1 min read

Open

On 2026-06-25 Corey Cooper announced a new USDC payment standard for the MPP…

Why it matters

On 2026-06-25 Corey Cooper announced a new USDC payment standard for the MPP ecosystem that leverages Circle Gateway and xReserve to make USDC chain-abstracted.

Key details

Developers can accept USDC from multiple chains while settling on a preferred destination chain, eliminating the common requirement that payers hold USDC on the merchant's settlement chain.
Agents can pay from wherever they hold USDC without bridging funds before each transaction, and blockchains can join MPP commerce without merchants and payers sharing the same settlement network.

Brief

Corey Cooper (2026-06-25) outlines a new USDC payment standard for the MPP ecosystem that leverages Circle Gateway and xReserve to make USDC chain‑abstracted. He argues this lets developers accept cross-chain USDC while settling on a chosen destination chain, allows agents to avoid per-transaction bridging, and enables blockchains to participate without shared settlement networks.

By @jerallaire

16 Twitter/X 2026-06-24 1 min read

Open

Brooke Hopkins, founder and CEO of Coval, announced a $28.2M Series A (reported…

Why it matters

Brooke Hopkins, founder and CEO of Coval, announced a $28.2M Series A (reported 2026-06-24); Coval processes tens of millions of calls per month for customers including Perplexity and Deepgram.

Key details

Hopkins argues voice is the killer AI interface and the first productionized use case for autonomous agents, claiming evaluation infrastructure she led at Waymo proved transferable to voice agents.
She narrowed a broad 'evals' idea into a focused enterprise simulation and observability platform; a customer offered to pay before she had written a single line of code. Key interview timestamps: 01:16 (voice as interface), 05:44 (missing infra), 12:32 (Waymo lessons), 16:38 (product-market fit), 25:35 (leaving Waymo).

Brief

Brooke Hopkins, founder and CEO of Coval, explains how her simulation and observability platform—used by Perplexity and Deepgram to process tens of millions of calls monthly—secured a $28.2M Series A (announced 2026-06-24). Drawing on Waymo evals infrastructure, she argues voice is the first productionized autonomous-agent interface and recounts a customer offering to pay before any code existed.

By @ycombinator

17 Twitter/X 2026-06-25 1 min read

Open

Joe Hudson, who coaches OpenAI's research team (including @sama) and leaders at…

Why it matters

Joe Hudson, who coaches OpenAI's research team (including @sama) and leaders at Apple and Google, published a 5-part framework on 2026-06-25 arguing 'emotional clarity' is the decisive skill for leaders in AI-forward environments.

Key details

Hudson claims AI excels at knowledge and effort, so human advantage comes from staying in hard conversations, not turning on oneself or others, pushing through failure, and building teams that 'metabolize fear' using concepts like the 'NBA-ification of teams' and a 'wisdom stack'.

Brief

Joe Hudson, an executive coach for OpenAI's research team and leaders at Apple and Google, outlines a five-part framework (published 2026-06-25) arguing that in AI-driven work the decisive advantage is 'emotional clarity.' He links concepts like the 'NBA-ification of teams,' a 'wisdom stack,' and methods to help teams 'metabolize fear.'

By @lennysan

18 Garry's List 2026-03-20 7 min read

Open

Yes, Fare Gates Actually Reduce Crime

Why it matters

BART installed reinforced fare gates at 48 of 50 stations in 2025; comparing the first seven months of 2024 vs 2025 violent incidents fell from 203 to 130 (−36%), overall crime fell 41%, robberies fell 71%, car burglaries fell 64%, and fare evasion fell 21%.

Key details

The new plexiglass gates produce roughly $10 million annually and have saved 961 hours of maintenance work; reporting also found ~80% of those arrested for crimes on BART had not paid fares, and these results came despite 21 officer vacancies.
The article parallels NYC’s 1990 Bratton/Jack Maple turnstile enforcement: investigators then found 1 in 6 fare evaders had outstanding warrants and 1 in 82 carried an illegal weapon; the author cites the 2024 Powell Street murder of Corazon Dandan by a previously banned fare evader as an outcome of non-enforcement.

Brief

BART’s 2025 rollout of reinforced fare gates at 48 stations produced a natural before‑and‑after test: in the first seven months after installation violent incidents dropped from 203 to 130 (−36%), total crime fell 41%, robberies −71%, car burglaries −64%, and fare evasion −21%. The gates also reportedly generate about $10 million a year and cut 961 maintenance hours, while police data show roughly 80% of arrested perpetrators were non‑payers; BART achieved these results with 21 officer vacancies. The piece links this to Bill Bratton and Jack Maple’s 1990 NYC strategy — the “turnstile as filter” — citing findings that 1 in 6 fare evaders had outstanding warrants and 1 in 82 carried weapons. The author argues the evidence undermines a pure “root causes only” approach and presents enforcement as a practicable, revenue‑offsetting public‑safety measure.

By Garry Tan

19 Twitter/X 2026-06-25 1 min read

Open

Author @istdrc asserts Slack is the bottleneck

Why it matters

Author @istdrc asserts Slack is the bottleneck: because Slack is built for humans you won't run ~10 agents per person there, so teams end up with one bot (e.g., Claude Tag).

Key details

Raft (@raft_hq) is designed for many named agents with roles in the same room as the team; the author claims 10 agents to 1 human is normal and agents hand off to each other.
Claude Tag (Claude, @claudeai) integrates into Slack as a team member with chosen channel and tool access; tagging it delegates tasks to Claude while the user focuses on other work (post dated 2026-06-25).

Brief

Claude Tag integrates into Slack as a team member with selectable channel and tool access so users can tag it to delegate tasks. @istdrc argues Slack's human-first design makes running ~10 agents per person impractical, and contrasts this with Raft (@raft_hq), which is built for many named agents that hand off work internally.

By @istdrc

20 Twitter/X 2026-06-25 1 min read

Open

The post claims the real value is in the 43 edits that were deleted immediately…

Why it matters

The post claims the real value is in the 43 edits that were deleted immediately after release; the author insists judgment resides in the diff between the first draft and the final deck, not in the polished output.

Key details

Kunal Tangri is quoted saying 30 years of 'capture your institutional knowledge' meant indexing and search, which strips judgment; @saranormous labeled the resulting gap the 'untrainable territory,' and @farsight_ai says it is building an instrument called the System of Judgment to live inside it.

Brief

@ginacostag_ argues the final deck hides value: 43 edits deleted immediately after publication show judgment lives in the diff between drafts rather than the polished output. Quoting Kunal Tangri, she says three decades of 'capture your institutional knowledge' reduced knowledge to searchable documents and discarded judgment; @saranormous named this the untrainable territory and @farsight_ai is building a 'System of Judgment.'

By @ginacostag_

21 Twitter/X 2026-06-25 1 min read

Open

Author @dr_cintas (published 2026-06-25) says Claude Tag launched 48 hours…

Why it matters

Author @dr_cintas (published 2026-06-25) says Claude Tag launched 48 hours earlier and calls the 'Slack-AI startups are over' reaction a recurring, usually-wrong hot take whenever a big lab enters a category.

Key details

The post argues winners 'out-care' rather than out-spend and promotes Viktor (viktor.com), claiming Viktor has worked on this problem longer than Anthropic, lives in Slack and Microsoft Teams, connects to 3,200+ tools, and offers $100 in credits with no card.
The post links a Fryd Wiatrowski thread titled 'Anthropic killed Viktor.com. A post-mortem,' which frames Anthropic's launch as directly competitive and includes a founder's response about opportunities from the launch.

Brief

Claude Tag's launch (noted 48 hours earlier in a 2026-06-25 post) prompted @dr_cintas to reject the predictable 'Slack-AI startups are over' hot take, arguing successful teams 'out-care' not out-spend. The post promotes Viktor (viktor.com) as a longer-standing Slack/Teams AI with 3,200+ integrations and $100 free credits, and links a Fryd Wiatrowski thread claiming Anthropic's launch targeted Viktor.

By @dr_cintas

22 Twitter/X 2026-06-25 1 min read

Open

Micron stock is up 20% on 2026-06-25 after record earnings, with the post stating…

Why it matters

Micron stock is up 20% on 2026-06-25 after record earnings, with the post stating AI-driven demand for memory now exceeds available supply.

Key details

The author predicts Apple will raise device prices—saying the next iPhone will cost significantly more—because memory costs have surged, and claims Apple previously forced Micron to sell memory “super cheap” during the last downturn.
Rolfe Winkler tweeted that Tim Cook blamed “the memory guys” for Apple raising prices, and a Micron executive Winkler interviewed pointed the finger back at Apple.

Brief

Micron's stock jumped 20% on June 25, 2026 after record earnings, which the author says proves AI demand for memory exceeds supply. The post warns Apple will raise device prices (including the next iPhone) because of rising memory costs and alleges Apple had earlier forced Micron to sell memory very cheaply; Rolfe Winkler highlighted mutual blame between Tim Cook and a Micron executive.

By @cryptopunk7213

23 Twitter/X 2026-06-25 1 min read

Open

swyx launched the first-ever music corner at AIE and invited coders who play…

Why it matters

swyx launched the first-ever music corner at AIE and invited coders who play music to impromptu jams between sessions, offering to sing and asking for participants' "best dad rock."

Key details

Sponsorship for the music corner plus the networking night on June 30 are the only remaining show-floor sponsorship opportunities at AIE; everything else is sold out.
Tickets for AIE are on track to sell out tomorrow (author's claim).

Brief

swyx announced he's pulled the trigger on the first-ever music corner at AIE, inviting developers who play music to impromptu jams and offering to sing—he specifically asked for participants' "best dad rock." Sponsorship spots for this music corner and the networking night on June 30 are the only show-floor opportunities left at AIE, and tickets are projected to sell out tomorrow.

By @swyx

24 Twitter/X 2026-06-25 1 min read

Open

@alex_prompter (2026-06-25) claims ~50% of current “AI automation” requires human…

Why it matters

@alex_prompter (2026-06-25) claims ~50% of current “AI automation” requires human monitoring—checking logs, restarting runs, and ensuring laptops stay awake, which he calls a new job with worse hours.

Key details

Hyperagent (@hyperagentapp) offers agents their own machine, browser, and code execution that keep running when you close your laptop; promotional credits: $100 at signup and $500 for migrating an OpenClaw setup.
Andrew Busse echoed the pitch, warning OpenClaw users about daily crashes, leaked secrets, and constant monitoring, and linked hyperagent.com/anonymous and a demo video.

Brief

@alex_prompter argues that roughly half of today’s “AI automation” still needs human babysitting—checking logs, restarting runs, and keeping laptops awake—so it isn’t true leverage. He points to Hyperagent (@hyperagentapp) as a fix: dedicated machine, browser, and code execution that continue running when you close your laptop, with $100 signup credits and $500 to migrate an OpenClaw setup; Andrew Busse amplified the message and linked a demo.

By @alex_prompter

25 Twitter/X 2026-06-25 1 min read

Open

On 2026-06-25 David Ch (@chhddavid) tweeted a Claude-based “Shipper” system that…

Why it matters

On 2026-06-25 David Ch (@chhddavid) tweeted a Claude-based “Shipper” system that scrapes every local business in a city in real time and filters prospects by review count, rating, last update date, and site quality.

Key details

The system pulls strongest photos and copy from a business's Google Maps listing, samples the brand palette, AI-rebuilds a brand-matched website plus a mobile app, writes a postcard quoting a real reviewer, and mails it to the owner by first name with a preview QR (the post also says it “emails the owner a postcard”).
Every step from discovery to brand-matching to outreach is automated; David invited replies of “SHIPPER” + RT and promised to randomly send free credits to try it out.

Brief

David Ch's 2026-06-25 post describes a Claude-based “Shipper” automation that finds local businesses with poor or no websites, filters them by review count/rating/last update/site quality, extracts photos/copy from Google Maps, samples brand colors, auto-builds a brand-matched website and mobile app, generates a postcard quoting a reviewer, and mails it to the owner with a QR preview; he offered free credits for replies of “SHIPPER” + RT.

By @chddaniel

26 Twitter/X 2026-06-25 1 min read

Open

Gina Costa (posted 2026-06-25 18

Why it matters

Gina Costa (posted 2026-06-25 18:27:16+00:00) stated: The hard part of video was never having ideas — it was turning them into something watchable.

Key details

OpenArt (@openart_ai) announced OpenArt Director and a new workflow called VIBE DIRECTING on 2026-06-25, claiming users can turn an idea into professional-quality video via a conversational 'vibe directing' process similar to vibe coding.

Brief

Gina Costa reposted OpenArt's June 25, 2026 announcement of OpenArt Director, a VIBE DIRECTING workflow that promises to convert an idea into professional-quality video through a conversation. OpenArt pitches vibe directing as the next step after vibe coding, saying users can chat ideas to perfection and walk away with watchable, unique films.

By @ginacostag_

27 Twitter/X 2026-06-25 1 min read

Open

@kimmonismus (2026-06-25) claims most AI agents produce drafts that are "80%…

Why it matters

@kimmonismus (2026-06-25) claims most AI agents produce drafts that are "80% there" and then require "an hour of cleanup" for customer-facing materials, failing on the last-mile.

Key details

Mutiny is described as purpose-built for go-to-market: it "pulls your CRM, your transcripts and your brand" and returns on-brand, ready-to-send customer-facing content.
Jaleh Rezaei (@jalehr) announced a launch saying it "blows up the entire GTM stack," replaces it with one agent that takes accounts "from cold to closed," and offered to "triple your AI credits" for anyone who comments "Mutiny."

Brief

@kimmonismus argues most AI agents handle busywork but stumble on the customer-facing last mile, leaving content "80% there" and needing "an hour of cleanup." He highlights Mutiny as a GTM-specific agent that pulls CRM, transcripts and brand to deliver ready-to-send materials; Jaleh Rezaei (@jalehr) claims the launch replaces the entire GTM stack and offers triple AI credits for comments.

By @kimmonismus

28 Twitter/X 2026-06-23 1 min read

Open

@jxmnop left their PhD at the end of 2025 and co-founded Engram

Why it matters

@jxmnop left their PhD at the end of 2025 and co-founded Engram

Key details

Engram builds per-user AI models that update approximately every minute and are designed to learn over time, remember details, adapt, and interact like a person
Engram positions itself among SF startups with the message 'scaling compute on your context', arguing current models don't deeply understand users' ongoing work

Brief

Author @jxmnop announces they left their PhD at the end of 2025 to co-found Engram. Engram is building per-user AI models that update about every minute, aiming to learn over time, remember user details, adapt and interact like a person, and promotes itself with the idea of scaling compute on user context.

By @jxmnop

29 Twitter/X 2026-06-25 1 min read

Open

On 2026-06-25 @johnloeber claimed OpenAI and Anthropic are rolling out new…

Why it matters

On 2026-06-25 @johnloeber claimed OpenAI and Anthropic are rolling out new frontier labs in a 'slow and staggered' way and that, since Mythos, we're at the 'beginning of the end' of public access to frontier intelligence.

Key details

He argues economic benefits will be internalized by the labs and a small set of trusted parties—likely large corporations and government—concentrating wealth while the public, already 'cartoonishly negative' about AI, sees few opportunities.
He predicts political backlash beyond current anti-data-center protests if the general public feels excluded, framing the situation as a tradeoff between safety and political/economic inequality and offering no prescriptions.

Brief

The Danger of Internalization: @johnloeber warns that OpenAI and Anthropic's 'slow and staggered' frontier lab rollouts (post-Mythos) will curtail public access to frontier models, concentrate economic gains among labs and trusted parties (likely corporations and government), and risk severe political backlash as public sentiment toward AI hardens; he offers no solutions.

By @johnloeber

30 Twitter/X 2026-06-25 1 min read

Open

Garry Tan called Legora “the defining legal AI startup”; Legora is a Y Combinator…

Why it matters

Garry Tan called Legora “the defining legal AI startup”; Legora is a Y Combinator Winter 2024 (W24) alum.

Key details

Cooley LLP launched Cooley GO Lab on 2026-06-25 using the WeAreLegora Portal to embed Cooley’s knowledge and workflows for YC founders, aiming to speed routine startup legal work.
Business Insider coverage by Melia Robin (link in the post) frames the partnership as Cooley urging founders to stop relying on chatbots for legal work; Legora’s press release is at legora.com/newsroom/cooley-l…

Brief

Legora, a YC W24 alum that Garry Tan called “the defining legal AI startup,” is powering Cooley LLP’s Cooley GO Lab via the WeAreLegora Portal. Launched 2026-06-25, the portal delivers Cooley’s workflows to YC founders to accelerate routine legal tasks and push founders away from using chatbots for legal advice, per Business Insider (Melia Robin).

By @garrytan

31 Twitter/X 2026-06-25 1 min read

Open

@fromzerotomill (2026-06-25) claims to leak the full 'Claude + FanPro' AI UGC…

Why it matters

@fromzerotomill (2026-06-25) claims to leak the full 'Claude + FanPro' AI UGC system: Claude reverse-engineers highest‑converting hooks and scripts and 'spins every angle'.

Key details

FanPro Studio converts those scripts into short clips in minutes, then the creator distributes across multiple accounts and relies on high volume to 'remove lucky dips' and optimize conversion.
The author says three things matter—'hooks that stop the scroll (visual + written)', 'volume that removes lucky dips', and 'content designed to convert, not just to get views'—and asks followers to Like + Comment 'AI' and follow to receive the system via DM.

Brief

fromzerotomill (25 Jun 2026) promotes a leaked 'Claude + FanPro' AI UGC workflow: use Claude to reverse‑engineer top‑converting hooks/scripts and generate angle variations, have FanPro Studio render those scripts into clips in minutes, then flood multiple accounts to prioritize conversion metrics over raw views. The post ends with a Like+Comment 'AI' CTA and follow requirement to get the system by DM.

By @fromzerotomill

32 Twitter/X 2026-06-25 1 min read

Open

@indexsy (published 2026-06-25) says there’s a one-time-fee app that lets you…

Why it matters

@indexsy (published 2026-06-25) says there’s a one-time-fee app that lets you bring your own API key while still hosting the app so you can share beautiful client reports; they’ll DM the link if you comment "TRACKER" + like and follow.

Key details

Charles Floate (@Charles_SEO) advises building a DIY SEO rank tracker in under 10 minutes: sign up for DataForSEO (or alternative), craft an exact prompt, use Codex or Claude Code to generate the tracker, host locally to save money, input your API keys, add pages to track, and refresh for updates.
Both posts promote avoiding 'extortionate fees' on rank trackers and gate the full prompt/link behind engagement (like + reply), using community distribution to share the build instructions.

Brief

@indexsy endorses Charles Floate and says a one‑time‑fee app lets you bring your own API key while still hosting shareable client reports; they’ll DM the link if you comment “TRACKER” and like (must follow). Floate’s method (DataForSEO, a tailored prompt, Codex/Claude Code, local hosting) builds a DIY rank tracker in under ten minutes.

By @indexsy