What AI topic do you want me to break down in detail next weekend? • Astra safety & containment • The SVP Chain agent-to-agent economy • Energy & data center bottlenecks • Open-weight models catching up Drop your pick in the comments 💬 #AgenticAI# #SVPChain# #AIResearch# #CryptoAI# #TechDiscussio#

The longer a video generation gets, the harder it is to know which past frames to remember. This paper picks them based on future usefulness instead of present similarity. Title: FrameMorrow: Future-guided Frame Selection with Prospective Tokens for Long-Horizon Video Generation URL: 📝 Overview FrameMorrow selects historical frames based on relevance to future information needs rather than similarity to the present, using "prospective tokens" that predict what will matter next without actually generating future video. ❗ Problem it solves Existing methods keep frames similar to recent content, but that's myopic — visually similar frames often don't help future generation, and long videos lose coherence as a result. 🛠️ Method A frozen visual encoder (DINOv2) feeds a small set of learnable projections and attention layers that autoregressively predict prospective tokens, trained via "future-grounded ranking distillation" against actual continuation frames. Because the generator itself stays frozen and FrameMorrow just hands it explicit selected frames through a per-backbone adapter, it plugs into open-source models as well as closed-source APIs like Seedance 2.0 and Kling O3. 📊 Results Tested across 5 benchmarks and 11 models, it lifts interactive-generation consistency from 84.95 to 89.30 and improves action alignment on the WorldMem model from 0.866 to 0.889. In a user study, raters preferred FrameMorrow's consistency 62.7% of the time versus 16.7% for native models. #VideoGeneration# #AIResearch#

⏳ The belief that "people used to be more moral" can fade after just 22 minutes talking to an AI that only knows text written before 1930. Title: Time Machine Experiments: Using Historically-Bounded AI for Inquiry into the Human Mind URL: 🧠 Overview This study tests whether the "illusion of moral decline" — the bias that past societies were more moral than today's — can be reduced by talking with Talkie, an AI trained only on pre-1930 text. 🎯 The problem it solves Archives can't answer new questions, and living witnesses have their memories distorted by knowledge of what came after. Until now there was no way to actually experience a "raw" historical perspective. 🛠 Method Instead of prompting a modern model to roleplay a historical persona (which leaks modern knowledge), the researchers used a model whose training data itself stops at 1930. 240 participants were randomly split between the 1930-AI and GPT-5.5 conditions and discussed morally contested sentence completions with their assigned model. 📊 Results The 100-year moral-decline score improved from -0.80 to +0.07 in the 1930-AI group (d=0.56), with almost no change in the control. Participants revised their view of the past downward rather than inflating the present, and also reported significantly deeper reflection. #BehavioralScience# #AIResearch#

🔥 @TencentHunyuan Hy4 preview just cited a Zhihu post with a surprisingly simple challenge to DeepSeek’s mHC: What if the doubly stochastic matrix is overcomplicating things—and Identity actually works better? Today, let’s revisit the cited post, “Your DeepSeek mHC May Not Need the ‘m’,” by Zhihu contributor 涮月亮的谪仙人. After training Qwen3 1.7B and 8B dense models from scratch on 150B tokens, the experiments found: 💡 Identity HC > mHC > mHC-lite > orthogonal mHC In other words: simply setting H_res = Identity beat the Sinkhorn-Knopp constrained version. Here’s why 👇 1️⃣ What does mHC actually change? Standard Transformers have one residual stream. Hyper-Connections (HC) expand it to multiple parallel streams: 🔹 H_pre reads from the streams 🔹 H_post writes back to them 🔹 H_res mixes information between them DeepSeek’s mHC constrains H_res to a doubly stochastic matrix via Sinkhorn-Knopp, helping preserve norms and stabilize propagation. But the experiments suggest a simpler question: Do we need H_res mixing at all? 2️⃣ mHC seems to learn something close to Identity anyway For a single layer, the learned H_res is already close to Identity: diagonal ≈ 0.96, off-diagonal ≈ 0.01 But multiply H_res across many layers, and it gradually collapses toward a uniform 0.25 matrix. So each layer may look almost like Identity, while their cumulative effect becomes uniform mixing. The simplest fix? Just set H_res = I. 3️⃣ Identity preserves stream semantics With Identity, each residual stream stays where it is: Stream 0 stays Stream 0. Stream 1 stays Stream 1. No repeated reshuffling, no cumulative mixing, and Iᴸ = I. H_pre and H_post also no longer need to track where each stream has been repeatedly moved—they simply learn where to read and where to write. 4️⃣ But cross-stream communication still happens Setting H_res = I does not isolate the streams. The projection that generates H_pre and H_post already sees all residual streams, making H_pre input-dependent. So information can still be dynamically aggregated across streams before Attention/MLP and written back afterward. H_res isn’t the only mechanism for cross-stream interaction. 5️⃣ Why might Sinkhorn hurt? Repeated products of positive doubly stochastic matrices tend toward uniform mixing. In the Qwen3-1.7B experiment, after 56 HC modules, the minimum singular value of the accumulated H_res product reached just: 9.2 × 10⁻¹⁸ By ~10 layers, the four streams were already approaching the same 0.25 uniform mixture. Sinkhorn also comes with extra cost: 20 iterations, backward recomputation, extra parameters, and approximation error. Identity has none of these—and preserves the residual signal exactly. 6️⃣ More sophisticated alternatives didn’t win either The experiments also tested mHC-lite, softmax-weighted convex combinations, and orthogonal variants using Cayley/Givens transforms. The observed ranking remained: Identity HC > mHC > mHC-lite > orthogonal mHC The simplest design won. 💡 The takeaway DeepSeek’s mHC uses sophisticated manifold constraints to stabilize Hyper-Connections. But these experiments suggest that H_res itself may not need to be learned or mixed at all. Sometimes the best manifold constraint is the most boring one: H_res = I. Or, as the original post puts it: Maybe DeepSeek’s mHC doesn’t need the “m.” 😆 👉 Read the full Zhihu post for the training curves, mathematical analysis, H_res visualizations, and implementation details: #DeepSeek# #mHC# #Hunyuan# #Tencent# #LLM# #Transformer# #AIResearch# #AI#

Ask an AI to write a paper and it will happily report favorable numbers for experiments it never ran — this "plausible fabrication" has been the real wall for research automation. Earlier automated-research systems poured effort into standing up dedicated agent infrastructure and forcing a single clean story to completion. But somewhere in the long generation, factual integrity would fray, and unsupported claims slipped straight into the paper. Behind the flashy results, the foundation of reliability was quietly missing. Spark-to-Paper flips the framing. Instead of a new agent platform, it implements paper generation as 13 composable skills running inside Claude Code. The key is separation: context-dependent judgment goes to the model, while citation checks and LaTeX compilation go to deterministic code. Experiments fix their metrics and table structure before results are observed, blocking post-hoc adaptation. And a "Self-Refutation Loop" — endlessly revising a hypothesis the evidence keeps rejecting — is capped at seven cycles, recording failure as failure and moving on to a different direction. The results speak clearly. Citation validity hits 99.5%, above the human-preprint average, and fabrication detection climbs from 14% single-pass to 92% with the full adversarial-review stack. At roughly $8.1 and 3.2 hours per manuscript, you get a paper complete with editable figures. Rather than dressing every attempt up as success, it centers evidence integrity and the acceptance that "not every research direction works." A thought-provoking blueprint for reliability in long-horizon AI. Spark-to-Paper: End-to-End Research Paper Generation as a Composable Skill #AIResearch# #AgentSkills#

AI systems now write research papers autonomously — yet phantom references, method-code misalignment, and unreproducible scores have become endemic failures undermining scientific integrity. Title: Science One Framework: A verifiable autonomous research framework via Chain-of-Evidence Google Cloud's Science One Framework treats verifiability as a first-class architectural constraint through the Chain-of-Evidence principle, solving the trustworthiness crisis in autonomous AI research at its root. 🔍 Highlight 1 — Chain-of-Evidence (CoE): two foundational principles Completeness: every claim carries a recorded evidence chain. Correctness: each chain genuinely supports its claim. These two principles eliminate phantom references entirely — baseline systems showed rates up to 21% — while achieving best-in-class method-code alignment across all evaluated systems. The key difference from prior work: evidence chains are constructed at claim-generation time, not retrofitted as a post-hoc check. 🏗 Highlight 2 — Three-module architecture Problem Investigator builds citation graphs from up to 100 full-text PDFs via Semantic Scholar API, grounding every reference in retrieved data rather than model memory. Discovery Engine explores parallel solution branches while keeping immutable records of all raw evaluator outputs. Paper Writer and Claim Verifier binds every factual claim to specific workspace artifacts and conservatively reconciles misalignments rather than deleting them — preserving scientific transparency. 🏆 Highlight 3 — MLE-Bench and Parameter-Golf results Across five Kaggle competitions covering medical imaging, fine-grained recognition, and 3D perception: two Gold Medals and two Silver Medals. Won the 3D Object Detection task where every baseline system failed completely. On Parameter-Golf — a live LLM training competition under strict hardware and file-size constraints — achieved state-of-the-art as of April 27, 2026, while baselines could not produce valid submissions at all. Rigor and capability don't trade off. Science One outperformed five state-of-the-art systems including AI Scientist v2, AutoResearchClaw, and DeepScientist, setting a new standard for verifiable autonomous research. #AIResearch# #AutonomousScience#

.@AFResearchLab & the Experimental Operations Unit flew the XQ-67A Autonomous Collaborative Platform over the Nevada Test & Training Range–validating the advanced sensors & autonomous architecture needed to extend our operational reach in contested environments. See how we are accelerating combat-ready uncrewed airpower 👇

AI research is changing as agents take on more of the experimentation loop. Instead of manually running every trial, our researchers increasingly use agents to propose changes, launch parallel experiments, evaluate results, and decide what to try next. Across our AI Functions research workloads, training jobs have grown 8x. At that scale, launch latency and GPU reliability become critical. Here's how we're using AI Runtime to keep up with the pace of agent-driven research.

AI Code Suggestions in Snowsight are now GA. Under the hood, our Snowflake AI Research team moved SQL autocomplete from a 30B-A3B MoE to a 4B dense model specialized for Snowflake SQL. • Acceptance: 17.80% → 26.35% • Latency: Down 71% The full research:

AI agents given 6 days and $3K produced two research papers, and both were rejected. The people who had spent months on those questions graded what the AI agent wrote. The failure was judgment. The main runs used Claude Opus 4.8 with extra-high reasoning on the OpenClaw scaffold, chosen after dry runs across OpenAI and Anthropic models, including an early pilot with GPT-5.3 Codex that could not handle the scaffold. Execution was never the problem. The agents ran hundreds of experiments, debugged crashing GPU pods, and compiled camera-ready LaTeX without a human touching anything. They were honest about it too, because the logs show marketable claims being retired in favor of negative results rather than any reward hacking. The failure was judgment. Round after round of automated reviews came back negative, but each response narrowed the claim and added a caveat instead of redesigning the experiment. Neither run noticed it was short on ideas rather than money, since both ended with over half of the $3K unspent. – arxiv. org/abs/2607.27191 Title: "Can AI agents conduct open-ended AI research? Early evidence from two case studies"