40% of TikTok’s Irish workforce face redundancy. 667 jobs out of 1778 to be cut. However, 321 new roles will be created under new ai restructure. Microsoft also announced this week it was cutting 4,800 jobs globally. Meta announced in May that 350 jobs in Dublin were at risk, on top of around 750 jobs at Covalen, a major Meta contractor. That’s about 2000+ jobs lost in total if we also include Diageo, Oracle, Genpact etc. Each will be entitled to new jobseeker scheme: €450 per week for first 13 weeks, €375 per week for next 13 weeks, and €300 per week for final 13 weeks. €14,625 per person over 39 weeks!

From 0 to ZK Concept Bites #7#: Verifiable Provenance Last time we covered the redundancy problem, how Ethereum's "everyone verifies everything" model creates the bottleneck that real-time proving solves. This time, a different ZK property: provenance. 🛡️ Provenance is the ability to prove where something came from and what happened to it along the way. The clearest live example is Brevis Vera. Modern flagship cameras (Sony, Leica, Nikon, Canon) ship with C2PA, a standard that cryptographically signs every photo at the moment of capture. The image carries a verifiable record of which device produced it. Any edit afterwards breaks that signature. Vera fixes the break. Each edit between capture and publication runs inside Pico zkVM, which generates a proof that the edit was legitimate. The result is an unbroken chain from camera sensor to whatever image you're looking at. Provenance shows up beyond images too. Anywhere you need a verifiable chain of custody, this is the property doing the work. We just added a section on this to Part 4 of the refreshed From 0 Knowledge to Zero Knowledge series. Dive deeper there 👇

💡This new paper introduces SFGA, a statistics-first gating architecture for cost-aware SFT data procurement. Instead of sending every case to an LLM judge, SFGA starts with low-cost blind measurements across three intrinsic quality axes: 🔹 Diversity 🔹 Utility 🔹 Redundancy Each is evaluated together with confidence intervals. Only when evidence is weak, borderline, or conflicting does the system escalate to an adjudicative debate between a buy-advocate and a reject-advocate, resolved by a presiding verdict. And even after escalation, SFGA does not blindly trust the LLM judge. It audits the process through advocate swapping, explicitly measuring negativity skew and positional bias. The contribution is not a new estimator or debate algorithm, but the architecture itself: ✅ Cheap statistics first ✅ Confidence gating ✅ Selective escalation ✅ Bias auditing For AI data markets, this direction matters: LLM judges should be used where they add the most value — on cases that are genuinely ambiguous, costly, and worth debating. Paper:

Secure execution requires adding friction where it matters most: before critical actions are executed. Four safeguards protocols can use: 1. Redundancy 2. Staged approvals 3. Timelocks 4. Veto rights

ℹ️ Since taking over mMEV, @RockawayX has launched their own dedicated market-neutral DeFi and structured yield strategy, mROX. To avoid redundancy, mMEV will be discontinued on May 10th. If you hold mMEV, please withdraw and reinvest in mROX for the same strategy exposure. Withdraw → Reinvest →

📚 A framework that fixes redundant flat-retrieval RAG by navigating a knowledge graph hierarchically. Title: LeanRAG URL: 📦 Overview LeanRAG is a graph-based RAG framework that combines knowledge graphs with hierarchical retrieval. It was accepted to AAAI 2026. 🔍 The problem it solves Naive retrieval just collects related chunks flatly, leading to overlapping, redundant content and ignoring structural links between documents. LeanRAG traverses semantically aggregated upper layers to produce well-grounded answers with less redundancy. 🛠 The method (5 stages) ・Semantic aggregation: cluster entities into summary nodes with adjacency relations ・Knowledge graph construction: connect entities and summaries in a multi-layer graph ・Hierarchical retrieval: anchor the query at fine-grained entities, then traverse upward to gather evidence ・Redundancy-aware synthesis: streamline overlapping paths ・Generation: an LLM answers from the gathered evidence 📊 Results ・About 46% lower retrieval redundancy versus flat retrieval ・On the Mix benchmark, LeanRAG 8.59 beats HiRAG 8.08, GraphRAG 7.87, and LightRAG 7.61 ・Win rates of 97.3% vs NaiveRAG, 78.1% vs GraphRAG, 81.2% vs LightRAG, and 100% vs FastGraphRAG A compelling option for anyone building a low-redundancy, knowledge-graph-grounded RAG. #RAG# #KnowledgeGraph#

🎬 Subject-driven T2V that keeps a reference subject's identity even as it shuttles across domains—real ⇄ fantasy. Title: DomainShuttle: Freeform Open Domain Subject-driven Text-to-video Generation URL: DomainShuttle reconciles subject fidelity with flexible style adaptation. Three highlights worth your attention. 🧬 Domain-MoT Video and reference image are processed in two independent branches; the reference branch uses Domain-aware AdaLN, modulated by time plus a domain attribute (real human / object / background / fantasy subject). Text cross-attention is frozen to preserve the base model's language guidance. 📐 Video-Reference DualRoPE Reference tokens get a separate RoPE space from video tokens for precise subject-level spatial control. Video starts its temporal index at 1, reference is fixed at 0, and multiple subjects (or multiple images of one subject) are organized via positional offsets. 🔗 Cross-Pair Consistent Loss Training uses two different reference sets at the same timestep, suppressing overfitting to single-frame redundancy and extracting the subject's intrinsic features—independent of irrelevant visual properties. Cross-domain subject consistency hits CD-Score 0.861, +18.7% over SOTA (Kling 1.6 is 0.725). A practical win for real⇄fantasy style transfer. #VideoGeneration# #GenerativeAI#

Galactic Civilization Is Forcing a Complete Revolution in Chip Design Earth’s energy limits make terawatt-scale AI impossible on the surface. That’s why Terafab is deploying massive orbital compute clusters powered by constant solar energy, while redesigning silicon from the ground up for vacuum, radiation, and fully autonomous off-world labor. This Drives Four Non-Negotiable Architectural Choices: >>Radiation-hardened D3 family 80% of output is destined for orbit. The D3 isn’t a commercial chip with hardening added later—it’s built from the transistor up with triple-modular redundancy, finFET structural tweaks, and advanced ECC to survive ~10¹⁵ cosmic ray hits per year at 99.999% uptime. >> High-temperature vacuum optimization No air, no liquid cooling. Heat must radiate into space. By deliberately engineering the D3 to run safely at much higher junction temperatures, radiator mass and size drop dramatically—delivering roughly 10× better FLOPS/watt in orbit than any ground-based system. >> Massive on-chip SRAM for edge autonomy (AI5/AI6) Mars propellant plants and orbital construction can’t rely on Earth comms. Optimus robots need real-time bipedal control and reasoning entirely on-device. That’s why half the accelerator area in the AI5/AI6 lines is dedicated to enormous SRAM—shattering the memory wall and boosting effective bandwidth by an order of magnitude. >> Recursive design-manufacturing loop At 100–200 billion chips per year, 12–18 month cycles are obsolete. Terafab integrates design, lithography, fabrication, and orbital simulation under one roof. Iterate overnight, test in radiation chambers, revise masks, and reprint compressing development from quarters to days. This isn’t incremental progress. Every transistor decision is subordinated to making humanity multi-planetary and eventually multi-stellar. The chips being taped out right now are the literal substrate for the next branch of human civilization.

Today we’re releasing Qwen-Scope 🔭, an open suite of sparse autoencoders for the Qwen model family. It turns SAE features into practical tools： 🎯 Inference — Steer model outputs by directly manipulating internal features, no prompt engineering needed 📂 Data — Classify & synthesize targeted data with minimal seed examples, boosting long-tail capabilities 🏋️ Training — Trace code-switching & repetitive generation back to their source, fix them at the root 📊 Evaluation — Analyze feature activation patterns to select smarter benchmarks and cut redundancy We hope the community uses Qwen-Scope to uncover new mechanisms inside Qwen models and build applications beyond what we explored.Excited to see what you build! 🚀 🔗🔗 Blog: HuggingFace: ModelScope: Technical Report: