Briefing

DR-Gym is an open-source Gymnasium-compatible environment that trains and evaluates demand-response from the electric utility perspective, addressing the missing feedback loop in offline smart-meter datasets. The work pairs physics-based building demand profiles with a regime-switching wholesale price model calibrated to real-world extreme events and a configurable multi-objective reward. Unlike device-level simulators, DR-Gym targets market-level utility decision-making and demonstrates realistic, learnable scenarios using baseline strategies.

U.S. electricity consumption is expected to surge—NEMA forecasts a 55% increase in annual demand by 2050 and a 300% rise in data center energy use over the next decade—raising pressure on grid planning and market structures. FERC Chair Laura Swett signaled governance concerns at PJM, saying the 13 member states plus D.C. have fundamentally different regulatory models and announcing a July 2026 conference to explore reforms. In the market, Fluence Energy struck master supply agreements with two major hyperscalers ahead of schedule, a development Jefferies’ Julien Dumoulin-Smith described as material to Fluence’s data center thesis. Policy and operations pieces highlight tradeoffs: an E3 study shows medium/heavy-duty fleet electrification in California and Georgia could modestly reduce residential rates by 2035 if grid investments are well-timed, while CAISO reports EDAM prices and transfer volumes are within expectations and credits battery storage as a growing Western-grid resource.

Rethinking energy resilience for AI infrastructure (DCD Rehlko session, published 11 May 2026) argues power is now the limiting factor for data center growth. It advocates integrated backup and continuous‑power architectures, fuel flexibility across natural gas, HVO, hydrogen, RNG and BESS, and explicit trade‑offs between uptime, efficiency, curtailment and emissions in site design and operations.

ASML — the Dutch firm spun out of Philips in 1984 — became the decisive winner in modern photolithography by combining modular manufacturing, deep customer partnerships, and sustained R&D to commercialize extreme ultraviolet (EUV) lithography. EUV uses 13.5‑nm light produced by vaporizing tiny tin droplets with a dual‑pulse laser; because materials absorb EUV, the system images via ultra‑precise reflective optics (no lenses) and shrinks mask patterns about four‑times onto 300‑mm wafers. The company’s technical path included immersion lithography with 193‑nm light and the TWINSCAN dual‑stage architecture (which eliminated idle measurement time), followed by decades of work on mirrors, light sources and debris control that only reached production viability around 2018. ASML’s rise depended on coordinated public‑private programs (the 1997 Extreme Ultraviolet LLC with >$270M invested), acquisitions (Silicon Valley Group in 2001; Cymer for ~$2.5B), and a 2012 co‑investment where Intel, TSMC and Samsung bought a 23% stake to sustain EUV commercialization. TSMC collaboration helped achieve ~500 wafers/day throughput targets; today EUV tools price >$120M, ASML’s market cap tops ~$400B, and its global supply chain (>5,000 suppliers, heavy European sourcing) supports its effective EUV monopoly while insulating it from single‑country risks.

Austin Energy, the municipally owned utility serving about 1 million people in the Austin area (≈580,000 accounts), is moving from early clean‑energy leadership into a new phase of operational planning to hit 100% carbon‑free generation as a share of load by 2035. The utility reported 65% carbon‑free supply in 2024 and is implementing its Resource, Generation and Climate Protection Plan to 2035. Stuart Reilly (General Manager) and Lisa Martin (COO) describe a hybrid portfolio approach: continued aggressive clean procurement where it hedges costs, combined with local dispatchable capacity, utility‑scale batteries, distributed batteries, transmission import upgrades and demand response to manage local reliability and market exposure inside ERCOT.

Key technical and market details: Austin retired ~725 MW of local gas capacity at Decker Creek (≈300 MW in 2020 and ≈425 MW in 2022), which contributed to the emergence of load‑zone price separation that undermines the hedging value of distant PPAs. Management gives the example of PPAs at roughly $50/MWh becoming ineffective when local nodal or zone prices spike (e.g., $850/MWh during critical ramps). To shrink that local premium the utility is investing in transmission (about $100 million per year in its capital improvement plan for five years) and adding local resources. Recent and in‑flight procurement includes a contracted 100 MW / 4‑hour battery with Jupiter Power, a ~40 MW distributed battery deal with Base Power (paired with a customer program offering a $500 upfront rebate plus ~$300/year performance payments), and authorization to negotiate an additional 100 MW / 2‑hour battery. Austin Energy estimates ~500 MW of credible near‑term new load on a ~3,000 MW peak system (mostly beneficial electrification and infrastructure rather than primarily data centers). Reilly and Martin emphasize the muni model’s community engagement—returning over $100 million to customers after Winter Storm Uri—and the need to communicate trade‑offs among reliability, affordability and environmental goals while using local dispatchable resources (including a proposed gas peaker for black‑start and long‑duration reliability) as a bridge to deeper decarbonization.

The eBook 'Redefining the data center for AI' (DCD Edge Infra & Inference Channel, published 13 May 2026 on datacenterdynamics.com) compiles insights from Schneider Electric and partners, citing an Innovation Summit estimate that 88% of companies pursue AI, and provides practical guidance on designing AI-ready data centres—power, cooling, density, efficiency and scalable architectures.

Cerebras’ wafer-scale strategy has become strategically timely because model providers and users are explicitly buying "fast tokens." The company’s WSE-3 (TSMC N5) places roughly half of wafer area into very fast SRAM (44 GB on-wafer) and delivers ~21 PB/s of on-chip memory bandwidth, enabling decode-style kernels with low arithmetic intensity to run at very high realized FLOPs. In SemiAnalysis roofline comparisons the wafer can, in theory, realize orders of magnitude more token interactivity than HBM GPUs for memory-bound decode workloads; OpenAI reported Spark variants achieving up to ~2,000 tokens/sec/user on Cerebras hardware. However, Cerebras’s sparse FLOP marketing (125 PFLOPS FP16 sparse) masks dense throughput (~15.6 PFLOPS assuming an 8:1 sparsity factor), and the wafer’s real constraints are SRAM capacity per wafer and off-wafer I/O.

Those constraints drive system and business trade-offs. Each CS-3 engine block pulls ~25 kW and requires bespoke power-delivery (12×3.3 kW supplies, 84 Vicor bricks) and custom liquid cooling; facility plumbing must support ~100 LPM per server versus ~1.5 LPM/kW for typical GB300 racks. Off-wafer bandwidth is only ~1.2 Tb/s (≈150 GB/s) via 12×100 GbE FPGAs, which means large models and large KV caches cannot be streamed efficiently — pipeline parallelism (layer-wise sharding) becomes the practical scaling strategy, at the cost of pipeline bubbles and multiplying per-inflight-microbatch KV storage requirements. SemiAnalysis’s telemetry (≈432k requests, ~80B tokens) indicates a P50 request of ~96.3k tokens and ~50% of sessions above 128k, underscoring why memory capacity and I/O matter. Commercially, OpenAI’s December 2025 MRA (750 MW committed; option to 2 GW), a $1B loan, and a ~33.45M-share warrant package (S-1 fair value $82.02/share) tie Cerebras’ near-term fate to a single large customer and create a path to rapid scale — but execution hinges on wafer supply ramp, datacenter capacity (750 MW by 2028), and architectural fixes (hybrid-bonded DRAM/photonic wafers) to address SRAM scaling and the wafer's “island” I/O geometry limitations.

Shanu Mathew (@ShanuMathew93) rejects the 'off-grid only' thesis for hyperscale AI, citing xAI's Colossus as grid‑tied with 300 MW TVA interconnect across two xAI‑funded substations and curtailment obligations. On-site turbines accelerated time‑to‑power; long‑term economics and operational complexity favor hybrid architectures combining grid interconnect, captive generation, storage, and demand‑response.

Shruti Mishra reporting from Baidu Create in Beijing says Robin Li presented Baidu as a full vertical AI stack — chip, cloud, model, app — all built in China. Announcements included 30,000+ Kunlun cards in production, a Kunlunxin IPO, ERNIE 5.1 with a 94% pre‑training cost cut, multiple agents (DuMate, Miaoda, Huiboxing), a 10B DAU agent forecast, and rapid OpenClaw adoption.

Based on the abstract (full text not consulted), the paper analyzes how routing decisions in Sparse Mixture-of-Experts form a geometric coupling between routers and experts: gradients for a routed token point along the same input direction in router and expert weights. The authors validate this in a 1B-parameter SMoE, show that auxiliary load-balancing disrupts the coupling (making router directions ≈3× more similar), and introduce a parameter-free online K-Means router (running-average centroids + cosine assignment) that minimizes load imbalance with only modest perplexity cost, suggesting geometric coupling underlies effective specialization. Published 2026-05-12 by Ahrac, Hochwald, and Geva (arXiv:2605.12476v1).

Oregon's Public Utilities Commission issued an order implementing last year’s law that forces data centers to pay their share of utility interconnection and growth costs. The order assigns growth-related generation/transmission costs to the customer classes that cause them, allows PUC-reviewed non‑emitting special contracts, requires queuing for available clean energy, and notes staff support—but limited capacity—for a flexible large-load tariff; the author says this shows what a traditionally regulated state can do and praises the 'utility of utilities.'

Attractor Models introduce a two-stage iterative-refinement architecture where a backbone proposes embeddings and an attractor module solves for a fixed point with gradients via implicit differentiation, which keeps training memory constant and allows adaptive iteration depth. Empirically (abstract only), they report up to 46.6% perplexity reduction and 19.7% downstream accuracy gains in pretraining—e.g., a 770M model beating a 1.3B Transformer trained on twice the tokens—and strong few-shot reasoning (27M params, ~1,000 examples: 91.4% Sudoku-Extreme, 93.1% Maze-Hard). The paper also describes 'equilibrium internalization,' where fixed-point training moves initial outputs near equilibrium so the solver can be removed at inference with little loss. Full text was not available here (abstract only).

KV-Fold introduces a training-free long-context inference protocol that treats the transformer's KV cache as an accumulator in a left fold, appending newly produced keys/values per chunk and forwarding the enlarged cache. The method yields a stable, precision-robust recurrence (plateau insensitive to 10,000× precision changes) and achieves 100% exact-match retrieval on Llama-3.1-8B across 152 trials (16K–128K tokens, depth ≤511). Based on the abstract (full text not provided).

ORCE introduces a decoupled, order-aware approach to verbalized confidence: answers are generated first and confidence is estimated conditioned on the fixed question–answer pair. Using a sampling-based surrogate of multiple completions and rank-based RL to encourage higher confidence for more likely-correct responses, the method improves calibration and failure prediction without substantially hurting answer accuracy, addressing interference from prior joint optimization techniques.

Multi-Stream LLMs (Su et al., 2026) argue that single-stream chat/instruction formats (e.g., instruction-tuned ChatGPT-style models) create a bottleneck preventing simultaneous reading, thinking and acting. They propose instruction-tuning for parallel computation streams—splitting roles into separate input/output channels so each forward pass concurrently consumes and emits tokens across streams. The preprint (37 pages) reports this architecture should enable concurrent action, boost efficiency via parallelization, and improve security and monitorability; code is provided on GitHub.

Nous Research released Token Superposition Training (TST) on 2026-05-13, claiming a 2–3× wall-clock speedup at matched FLOPs without changing architecture, optimizer, tokenizer, or data. TST trains on averaged contiguous token 'bags' for the first third of training with a modified cross-entropy, then switches to standard next-token training; validated at 270M, 600M, 3B dense and 10B-A1B MoE.

The paper proposes Self-Supervised Laplace Approximation (SSLA), which sidesteps parameter posteriors and directly approximates the posterior predictive by refitting on model self-predictions; a faster approximate variant (ASSLA) avoids costly refits. The authors provide theoretical analysis and experiments on Bayesian linear models through Bayesian neural networks, showing better predictive calibration than classical Laplace methods while keeping computation efficient.

Foxconn, the world’s largest contract electronics manufacturer, reported $261 billion in annual revenue and employs more than 800,000 people; at scale it produced up to 5.8 billion items in 2021 and is estimated to control roughly 30–40% of the global contract-manufacturing market. Long dependent on Apple—its most important customer for over a decade because of the iPhone—Foxconn also builds devices for Amazon, Huawei, Sony, and Dell. Beginning in 2023 the company became Nvidia’s primary electronics partner, turning Nvidia’s chips into finished server products that populate AI data centers; that AI-hardware boom has given Foxconn a noticeable boost. Nevertheless, the article notes that Foxconn’s underlying contract-manufacturing business model continues to face structural strains despite the Nvidia-driven uplift.

NatPurser warns backlash to data centers is driven as much by transmission buildout impacts—buyouts, easements and eminent domain—as by the facilities themselves. While transmission upgrades are needed to replace aging lines and support electrification, the politics won’t separate lines serving data centers from lines serving communities, so resistance to transmission projects (e.g., Georgia Power seizing homes for a $17B data center) will broaden.

Aniruddh Mohan highlights a 2026-05-12 ESIG/Brattle paper (with a clarifying graphic) showing that large new electricity loads do not automatically raise or lower retail rates. Impacts depend on system-specific factors; Mohan warns unhedged utilities can pass higher costs to customers (notably in PJM), while Matthew Yglesias notes reduced prices are possible but not guaranteed.

VECA (Visual Elastic Core Attention) tackles the quadratic cost of ViT self-attention by introducing a small set of learned core embeddings (C) that mediate communication for N patch tokens, producing linear O(N) complexity for fixed C. Cores are initialized and propagated across layers; nested core-axis training permits elastic compute–accuracy trade-offs. The authors claim competitive results on classification and dense tasks with lower compute (see arXiv and GitHub).

Reward hacking in rubric-based reinforcement learning: Mahmoud et al. (arXiv 2026-05-12) analyze divergence between training verifiers and a three-judge reference panel across medical and science domains. They separate verifier failure from rubric-design limitations, show weak verifiers yield nontransferable gains and recurring exploitation, and propose the self-internalization gap (policy log-probabilities). Stronger verification reduces but does not guarantee broader quality gains. (Summary based on abstract only.)

Brad DeLong argues that headline and core inflation readings in mid‑2026 replicate the early‑2021 to mid‑2022 inflation pattern, but with a narrower, geopolitically driven center. He highlights specific CPI moves: core CPI at a 4.8% annualized April‑over‑March rate, headline CPI and shelter/housing at 7.2% annualized, energy +3.8% month‑over‑month and +17.9% year‑over‑year, and gasoline +28.4% y/y. Unlike the broad, overlapping shocks of 2021–22 (fiscal stimulus, supply‑chain bottlenecks, housing dynamics, and the Ukraine shock), today’s episode is primarily an oil/energy relative‑price shock that can still feed into wage and margin behavior. DeLong counsels the Fed to refrain from immediate hikes, extend the horizon for 'higher for longer' guidance into 2027, prepare models for possible tightening beginning late 2026, and focus on anchoring expectations to avoid costly re‑anchoring if wage‑price feedbacks emerge.

Lenny's May 11, 2026 newsletter highlights two practical playbooks for scaling AI inside companies. Sendbird CEO John Kim outlines an internal product called Automators — a gamified marketplace where anyone can post a quest (with risk, weeks‑saved, and beneficiaries) and engineers or AI agents claim work. Contributors earn experience points they can redeem for gift cards or executive time, turning AI adoption into a measurable, crowd‑sourced product rather than a top‑down program. Kim stresses secure, InfoSec‑vetted app templates so marketers and CSMs can ship production features (his marketing team launched a Stripe‑backed swag store in days). He also tracks org token consumption with tiers from <1M tokens/day to >100M tokens/day, watches the smoothness of token‑usage curves to ensure AI coverage around the clock, and formed a weekly cross‑functional AI task force plus an AI Engineer for Internal Operations to remove barriers; hiring criteria now emphasize curiosity, agency, and energy. On the engineering side, Notion’s Ryan Nystrom advocates spec‑first development for agentic coding: maintain human‑readable Markdown specs as source of truth, 'yap your spec' with Whisper transcripts, and point coding agents (Codex/Boxy) at specs to produce PRs with UI previews in ~20 minutes. He argues fast CI is the bottleneck on agent velocity — reducing CI time (target: to 25% of current) multiplies iteration throughput — and that good developer experience (CLI, docs, fast CI) benefits both humans and agents. Both guests converge on: enable non‑engineers safely, measure and celebrate AI fluency, and invest in developer workflows to amplify agent productivity.

The article tests Qwen3.6 27B under five quantization recipes (FP8 block128; AutoRound INT4 group128; AWQ INT4 group32 with MTP→BF16 and linear-attn ignored; NVFP4 via llm-compressor; and AutoRound NVFP4 with 4-bit local activations but linear-attn in 16-bit). Methodology measured accuracy across benchmarks (including coding pass@k/LiveCodeBench), latency, memory, and MTP tensor handling; large Linear weights were the primary quantization targets while embeddings, output head, norms and small state tensors were kept at higher precision. Results: full NVFP4 that quantizes linear-attention performs substantially worse across tasks, NVFP4 that preserves linear-attn in BF16 still slightly trails, and Intel’s INT4 recipe is surprisingly robust when key linear-attention parts (inproja/inprojb) are kept higher precision. The author also observed quantized variants produce more tokens and that higher token counts often coincide with lower accuracy; experiments ran on Verda B200 and RTX Pro 6000 GPUs.

The AI is 5% of the work. @alex_prompter argues that the remaining 95% — observability (Langfuse, Braintrust, Helicone), evals, durable runtimes (Temporal, Inngest), guardrails, memory (Pinecone, pgvector, Turbopuffer), tools (MCP servers, E2B, Modal, Browserbase), auth/multi-tenancy, cost controls, human‑in‑the‑loop, prompt versioning, orchestration, and model routing — is the real product. A CTO: observability + evals + durable runtime + guardrails = minimum viable production stack.

Larry Fink contends AI is not a bubble but undersupplied: power, compute, chips, and memory are the main bottlenecks, and token usage has surged in the last few months in America. He predicts exponential global demand growth and says governments won’t fund needed spend, so private institutional capital (notably BlackRock) must step in.

Anthropic’s Claude Mythos Preview (via Project Glasswing) passed two independent evaluations (UK AISI and XBOW) that reported breakthrough autonomous offensive-security capabilities: AISI says Mythos solved both end-to-end ranges including the never-before-cleared Cooling Tower and met a 2.5M-token cap; XBOW flagged unprecedented token-level precision. Anthropic is distributing Mythos to defenders while building safeguards and disclosure workflows.

Pion introduces an orthogonal equivalence transformation optimizer that modulates weight-matrix geometry by applying left/right orthogonal updates while strictly preserving singular values and the spectral norm. The paper derives the Pion update rule, studies design options and convergence properties, and presents empirical evidence that Pion is a stable, competitive alternative to additive optimizers (e.g., Adam, Muon) for LLM pretraining and fine-tuning.

Copper is nearing record highs after prices breached $14,000/t as rebounding Chinese demand meets tightening supply. Middle East disruptions have sparked a global sulfuric acid shortage—key for ~20% of copper output—pushing acid prices nearly double and Chinese inventories down. Sprott forecasts strategic uses will rise from 32% to 45% of demand over the next 15 years (by 2040).

The summit agenda lists eight items Trump and Xi will tackle, with priority on trade (extend the one-year tariff truce from the Busan Summit, push China to buy soybeans and Boeing), AI/chips and rare-earth leverage (US ban on Nvidia H200, Jensen Huang present, Chinese export controls), and urgent security issues — Feb 2026 US‑Israel strikes on Iran, a Strait of Hormuz blockade that sent fuel prices up, and Trump’s pledge to discuss Taiwan arms sales with Xi.

Mario Nawfal tweeted on 2026-05-13 that Boom Supersonic built a 42‑MW "Superpower" turbine in a shipping container, which runs in extreme heat without water cooling and is portable; he also highlighted China testing rail-free, virtually guided trams carrying 500 passengers at 70 km/h and urged U.S. leaders, naming Trump, to invest in futuristic infrastructure.

SenseNova-U1 introduces NEO-unify, a unified multimodal paradigm that treats understanding and generation as a single process. The authors release two models (8B dense and 30B MoE A3B) and claim parity with top-tier understanding-only VLMs across perception, reasoning, decision-making, and spatial tasks, while also achieving strong any-to-image synthesis and interleaved multimodal generation; full design and training details are provided on the project page.

LongMemEval-V2 (LME-V2) is a benchmark for assessing whether memory systems let agents acquire environment-specific experience; it provides 451 questions across five memory abilities with histories up to 500 trajectories (115M tokens). The authors evaluate a RAG-style AgentRunbook-R and a file+coding-agent AgentRunbook-C, reporting 72.5% accuracy for AgentRunbook-C versus 48.5% for the best RAG baseline and 69.3% for a coding-agent baseline, while noting higher latency; summary based on the abstract.

The Utility Dive Studio report (published 2026-05-13) synthesizes views from 135 energy executives on balancing growth, cost, and sustainability amid accelerating electrification and looming decarbonization targets. Respondents describe managing volatility, redirecting investment priorities, and maintaining strong confidence in the energy transition; the analysis is custom sponsored content by Utility Dive’s Studio.

Anthony Sagami (@anthonysagami) shared a WSJ article on 2026-05-13 describing orbital data centers: swarms of AI-chip satellites with solar arrays in near-polar orbits to maximize sunlight. He appended a direct investment pitch, listing 49 AI-related tickers (e.g., $NVDA, $MSFT, $META, $GOOGL, $AMZN) and urging buys.

Fred Stafford warns that framing transmission projects tied to data centers as “corporate land grabs” — plus blanket partisan rejection from MAGA — creates political barriers to decarbonization and reindustrialization, so he advocates prioritizing energy technologies that minimize land‑use conflicts. Nat Purser notes that anger over buyouts, easements and eminent domain for transmission will broaden opposition even though grid expansion is required for electrification.

SI-Diff addresses contact‑rich assembly (peg‑in‑hole) by learning a single force‑domain diffusion policy with a mode‑conditioning mechanism and a search teacher policy that generates diverse trajectories; it maps tactile and end‑effector velocity observations to actions. Experiments report extending x–y misalignment tolerance from 2 mm to 5 mm versus TacDiffusion and strong zero‑shot transfer to unseen shapes. Abstract only.

GlobalWafers reported Q1 2026 results on May 8, 2026 showing revenue of NT$13.98 billion, a 3.6% sequential and 10.3% year‑over‑year decline, with gross margin compressed to 20.8% (down 4.9 percentage points QoQ), operating margin at 10.5%, and EPS of NT$3.97 (vs NT$3.05 in Q1 2025). Chairperson Doris Hsu called Q1 “a relatively low point in the current cycle,” while SEMI shipment data released April 29 show wafer shipments are recovering (up 13% YoY in Q1 and positive YoY for seven straight quarters after a prior seven‑quarter decline). The company’s current margin profile remains far below its Q3 2022 peak (43.7% GM on ~NT$18.6 billion), and the author notes GlobalWafers has stopped publishing its revenue/gross‑margin chart after Q2 2025, reflecting continued but slow cyclical recovery and some non‑recurring near‑term margin headwinds.

Robert Friedland reports a Sept 2026 15,500–17,000 call spread of 5,500 lots traded through an LME broker on 13 May 2026 for about $22 million in premium. He interprets the block as institutional conviction — 'extremely bullish' and confirmation of a copper supercycle — and highlights Ivanhoe Mines' Kamoa‑Kakula operation.

Luis Garicano, endorsed by @patrickc, argues that Europe is falling behind the United States because commonly cited PPP comparisons at current prices mask real-volume gains in U.S. technology measured by constant-price PPP. He claims tech-driven productivity growth matters for wages and wealth because much consumption is non-tradable, tech firms bid up local wages, and a large share of tech gains accrues as high-margin profits and equity value rather than visible price declines. Garicano cites concrete data: Apple/Microsoft/Nvidia/Alphabet/Meta/Amazon market cap ≈$21 trillion, median Meta pay $388,000 (2025), and OECD (2021) median disposable incomes where Americans earn 30–52% more than comparable Europeans. He also cites a 2026 NBER paper showing the U.S.–Europe hours gap has narrowed and provides anecdotal evidence of a U.S. construction boom and higher entry wages, arguing these patterns make divergence self‑reinforcing and politically salient for European reform.

Fred Stafford finds the Fortune article interesting but condemns More Perfect Union's 'nearly 50,000' framing as 'the epitome of slopulism,' comparing it to past anti‑nuclear environmental reactions. More Perfect Union says NV Energy will stop servicing homes in the Lake Tahoe area next year to redirect electricity to data centers; NV Energy is ending a wholesale contract with Liberty Utilities in May 2027, while Liberty is seeking replacement power and denies immediate shutoffs.

Corey Ganim outlines a repeatable serviceized product — the $999 AI Assessment — that maps AI tools onto an owner’s existing workflows and acts as a feeder for $1k–$10k implementation projects. The workflow is: a 45‑minute Fathom‑recorded Google Meet to surface pain points; transcript analysis with Claude (he uses chained Skills) to identify 5–7 opportunities; a Gamma template report (executive summary, impact/effort matrix, 3–7 tool recommendations, 4‑day quick wins, financial impact) produced in minutes; and a 30‑minute review call that converts ~40–60% into paid implementations. He guarantees 5 hours/week saved or a full refund, cites an average of 6+ hours/week saved for ~$40/month in tools, and gives an example where a $3k custom GPT removed 95% of email work per listing. His go‑to‑market advice: learn the stack (Claude, ChatGPT, Fathom, Gamma, Zapier), do 2–3 free assessments, then scale pricing from $499 up to $999 and build an upsell ladder for the real revenue.

Nate (May 12, 2026) outlines how agentic commerce—software agents holding wallets, signing authorizations, and paying merchants—unbundles the single human checkout click into six commercial responsibilities that must be explicitly assigned: identity, authorization/evidence, fraud/risk, payment credentials, settlement/refunds/liability, and data/merchant–customer relationships. He traces the opening clash to OpenAI + Stripe’s Instant Checkout (shipped in September and scaled back five months later) and contrasts it with Shopify and Google’s counter-protocols, arguing the real fight is over where commercial trust (the persistent authorization/evidence layer) lives. Key technical claims: authorization is distinct from payment and must outlive the transaction; software-to-software payments (including stablecoin rails) should use different infrastructure; and AWS’s platform position matters. Nate ends with two operational prompts — a responsibility-layer audit and an authorization spec — for builders, finance, and legal to close gaps.

Rising gasoline prices — roughly a 60% increase in the U.S. since earlier this year — are making return-to-office mandates feel like pay cuts for many commuters: the average American travels about 26 miles to and from work, amplifying cost and time burdens. Governments from India to Indonesia, Myanmar, Malaysia, Sri Lanka and Vietnam are implementing measures to reduce commuting and gasoline demand. On the cybersecurity front, Anthropic’s Mythos is effectively constrained by U.S. policy, while OpenAI’s GPT-5.4‑Cyber and GPT-5.5‑Cyber (reported as strong) are being discussed with EU officials for access; several European AI labs (Mistral, H Company, LeCun’s team, Black Forest Labs, SAP) are racing to produce cyber-capable models. In markets, CoreWeave beat revenue expectations in Q1 2026 with $2.08B but reported a wider loss of $1.12/sh, guided Q2 to $2.45–$2.60B (below consensus), showed operating margin deterioration (‑7% vs. ‑3% YoY), and signaled heavy debt service (~25% of revenue) even as it scales power to 3.5GW under contract.

Dorialexander (2026-05-13) argues that when Freon, Kaon, random or inverted spectra produce similar results, the core training signal is the data-induced gradient subspace combined with boundary-conditioned iterations on the learned manifold. In the Deep Manifold/Dataualism framing, orthogonalization is a helpful human prior but can mislead because optimization trajectories shift; Kaon perturbs spectral magnitudes while keeping subspace orientation, and the author is testing these claims on SYNTH.

Qualcomm is at the center of skepticism in the current chip rally: while the broader semiconductor index climbed nearly 60% from early April 2026, Qualcomm’s stock has surged over 80% since the end of March largely on hopes the company can become a major data‑center AI player. That optimism persisted even after mixed late‑April results and guidance (current‑quarter midpoint revenue $9.6B and EPS $2.20 vs. consensus $10.3B and $2.43). A vague management remark about a custom hyperscaler chip deal — with specifics deferred to an analyst day in June 2026 — and CEO Cristiano Amon’s comments about AI wearables have amplified the narrative. Key Context warns the core mobile franchise is eroding (Apple shifting to internal modems; Android demand hit by rising memory prices), meaning the stock’s run may be priced for a transformation that is unproven and execution‑dependent.

CJ Handmer (2026-05-13) argues that offsetting old-age senescence and insecurity requires sustained growth of roughly 2–9% annually—driven by many young, ambitious people—not housing-driven, tax-advantaged intergenerational transfers. He cites Japan's low fertility and stagnant economy as a caution, proposes Australian super invested in manufacturing, but concludes only slowed aging or AI-driven hypergrowth can reliably escape the demographic trap.

NVIDIA is valued at $5.33 trillion, exceeding the combined GDPs of Germany, Japan, France, the UK, Italy, Canada and Brazil, the author warns. Citing the 2000 Microsoft episode (≈$600B then −60% in 18 months), the post argues NVIDIA's valuation is a speculative bet on chips used to train AI models that have not yet proven profitable and that there is no historical playbook if AI spending decelerates.

Discounted exposure to copper clad laminates (CCL) is presented as a value-investor play by Collyer Bridge in a May 12, 2026 Substack preview: the author argues investors can gain CCL exposure cheaply by owning a holding company that majority-owns a CCL producer while the holdco’s four other divisions account for nearly the entire market capitalization, effectively leaving the CCL business under‑valued. The memo cites acute CCL tightness and AI-driven demand as catalysts, pointing to a May 3, 2026 tweet from @jukan05 reporting a Seoul PCB maker placed advance orders worth 10 billion won with Taiwanese producers EMC and TUC (order volumes described as >5x typical), and frames the thesis as scuttlebutt-driven value research — full details and on‑the‑ground checks are behind the post’s paywall.

Garry Tan's episode on Rick Rubin's Tetragrammaton podcast (published 2026-05-13) runs about two hours and argues that engineers who resist 'vibe coding' and AI would benefit from adopting them. He recounts his path from computing, mathematics, and video games to engineering, YC founder selection, AI-driven creativity, builder intuition, institutional reinvention, and AI-enabled manufacturing.

Author @BoringBiz_ argues that AI value will concentrate at the application layer: apps require little capex vs. the model layer’s “billions” for GPUs and data centers, are asset‑light with lower leverage, can focus on vertical customers, and face less top‑heavy competition (OpenAI, Anthropic, Google), yielding a larger TAM.

Ben Halligan argues that Jack Dorsey’s approach — which he dubs “Dorsey Mode” — is a radical organizational playbook shift away from Andy Grove’s High Output Management, and that it has measurable early upside (Halligan says Dorsey’s first quarter after adoption was “a banger”). The model accelerates iteration so planning cycles break down, turning many 1‑way doors into reversible choices and elevating distribution as the primary moat. Recruiting now emphasizes AI problem cases or live demos, with some firms favoring very senior engineers (Meta, HubSpot) and others hiring junior, AI‑native talent. Org charts move from large triangular hierarchies to small teams orbiting a central world model; titles are being removed; decisions increasingly hand off to systems; IT becomes the scaffolding team that records and feeds meetings and context into models; compensation and leadership practices must change, and CEOs are expected to lead by building and running hackathons to push adoption.

DexTwist introduces a functional twist-retargeting method for MR-based teleoperation that targets contact-rich rotational manipulation where kinematic imitation fails. The system detects tripod pinches, estimates intended screw axis and twist, then performs a real-time joint-space residual optimization minimizing a virtual-object objective (turn angle, axis consistency, fingertip closure, tripod stability). Simulations and real tests demonstrate improved turning-angle tracking and reduced screw-axis drift versus a vector-based baseline.

CUActSpot targets the long-tail of complex GUI interactions that undermine computer-use agents by providing a multimodal benchmark (GUI, text, table, canvas, natural image) and a renderer-based data-synthesis pipeline: automatic scene generation, screenshot/element-coordinate recording, and LLM-produced instructions/action traces. Training on this corpus yields Phi-Ground-Any-4B, which outperforms open models under 32B parameters. Only the abstract was available for this summary.

In a May 11, 2026 interview with Ryan Peterman, Harvard CS professor David J. Malan — the instructor behind CS50 — explains why his lectures resonate, how AI is reshaping student behavior, and why C still matters. Pedagogically, Malan builds a single “memorable moment” into each lecture (his classic is ripping a phone book to anchor binary search) and projects high energy driven by a fear of boring students. On AI’s downstream effects he reports fewer students enrolling and fewer companies hiring junior engineers from Harvard; cheating still affects roughly 5–10% of students per semester, but AI-produced responses are difficult to tie to a source so prosecution is harder. Pedagogically he defends teaching C because its small, low-level surface forces students to implement basic data structures and exposes machine-level concepts; he names pointers as the hardest topic for novices. The full conversation and transcript are linked on Substack and major podcast platforms.

Paul Graham argued at YC | Stockholm (April 29, 2026) that founders seeking faster capital, serendipitous meetings, and industry respect should relocate to Silicon Valley, citing faster-moving investors, a pay-it-forward culture, and the Dropbox story. He also outlined how to help Stockholm thrive, recommended YC as the optimal path, and asked if Stockholm could become Europe’s Silicon Valley.

A popular dining-focused credit card received an anniversary refresh on May 11, 2026 that adds a 5X hotel-earning rate, nearly $100 in limited-time travel and dining credits, and rental-car elite status (enrollable now). AwardWallet notes the upgrade is coupled with a welcome bonus up to 100,000 points, though the email withheld the issuer/card name for advertising reasons.

Task-Adaptive Embedding Refinement via Test-time LLM Guidance presents a test-time method that refines query embeddings via a generative LLM's feedback on a small document subset to tailor embeddings to ad-hoc zero-shot search and classification tasks. Experiments with state-of-the-art embedding models across diverse benchmarks report consistent gains (up to +25% relative), improving ranking and class separation; code released on GitHub. Abstract only; full text not provided here.

The paper introduces a proximal-gradient Monte Carlo sampler for composite log-concave densities π ∝ e^{-f-g}, combining gradient steps on smooth f with a restricted Gaussian-oracle (RGO) proximal sampler for g. Under α-strong convexity of f+g and β-smoothness of f it achieves ε total-variation error in ~O(κ·sqrt(d)·log^4(1/ε)) iterations (κ=β/α), and the authors extend guarantees to Poincaré/LSI targets and to Lipschitz non-smooth f.

Ramp's merch fulfillment service (announced in an email published 2026-05-11) combines printing with inventory storage, pick-and-pack operations and worldwide shipping, plus online store setup, onboarding kits and event/campaign distribution. The offering is modular — clients can select full fulfillment or specific services — and was credited with selling out The Bugle Podcast's limited-edition Christmas jumpers after Ramp printed and shipped orders globally.

Fibre‑optic, wire‑guided FPV drones are reappearing as a transformational weapon: by tethering control and video over fibre they are effectively immune to RF jamming, offer much lower latency and higher bandwidth for sharper targeting, and have become decisive on Ukraine’s battlefields. The Economist (May 11, 2026) reports a ten‑page GRU proposal to supply Iran with 5,000 fibre‑optic drones, long‑range Starlink‑guided UAVs and training, including plans to attack US landing forces. FPV ranges are now approaching ~40 km, and CNAS modelling and US Indo‑Pacific Command concepts (eg “Hellscape”) show dense layered FPV attacks (80 km→5 km) would threaten amphibious operations such as a Chinese invasion of Taiwan. The piece warns of accelerating tech flows between Russia, China, Iran and North Korea, wider proliferation to proxies (eg Hizbullah), and related operational impacts including quieter Chinese submarines and recent Iranian strikes (May 7) that reportedly damaged hundreds of US‑linked facilities.

MoLA (Mixture of Latent Actions) targets the gap between video-based imagination and actionable control: instead of feeding predicted frames to a policy or decoding videos directly into controls, it infers a mixture of latent actions via pretrained, modality-aware inverse-dynamics models (semantic, depth, flow) to produce a physically grounded action interface. Evaluated on LIBERO, CALVIN, LIBERO-Plus and real robots, the abstract reports consistent improvements in success rates, temporal consistency, and generalization; summary based on the abstract (full paper not reviewed).

Gary Marcus amplifies Noam Brown’s claim that, for current AI systems, intelligence scales with inference compute and should be measured as intelligence per token or per dollar (model comparisons became unreliable in 2024). Marcus adds that humans achieve high intelligence on ~20 watts, so new architectures may rival raw compute gains in the long term.

EgoEV-HandPose tackles egocentric 3D bimanual hand-pose estimation and gesture recognition from stereo event cameras by introducing KeypointBEV, which lifts stereo features into a bird's-eye-view and iteratively reprojection-refines depth and kinematic estimates. Trained and evaluated on the new EgoEVHands dataset (5,419 sequences, 38 gestures), it reports MPJPE 30.54 mm and 86.87% Top-1 accuracy, outperforming RGB-stereo and prior event-based methods, notably under low-light and occlusion.

The Economist's World in Brief surveys mounting political, economic and supply‑chain stresses: Labour’s leadership is in crisis after a local‑election rout — more than 80 MPs sought Keir Starmer’s resignation, Miatta Fahnbulleh and four junior ministers resigned, and 30‑year gilt yields hit levels not seen since 1998. Geopolitical tensions with Iran have sent Brent to about $105/barrel and complicated ceasefire talks, while Washington warns of fragile diplomacy. Domestically in the US, consumer‑price risks loom: food prices are roughly 33% above pre‑pandemic levels, April CPI data are imminent, tomatoes are ~25% more expensive year‑on‑year, and higher fertiliser, fuel, plastic and transport costs (transport accounts for up to half of supply‑chain oil use) threaten further grocery inflation. Meanwhile, defence spending and industrial disputes are heating up: a proposed ~40% jump to $1.5trn Pentagon outlays for 2027 and a potential Samsung strike affecting ~40,000 workers could both have global economic impact.

Laplacian-based neural operators are analyzed for the generalized Gierer–Meinhardt reaction–diffusion system: the paper obtains explicit approximation-error bounds depending on network depth, width, and spectral rank by using the Laplacian eigenfunction expansion of the PDE Green’s function. The authors show parameter complexity scales at most polynomially with accuracy and present numerical experiments consistent with theory. Summary based on the abstract; full text not reviewed.

Yglesias argues on May 12, 2026 that the practical project of winning voters who hold a mishmash of views should be reframed as “shmoderation” rather than the staid label “moderate.” He recommends a problem‑solving, eclectic political brand exemplified by the House Problem Solvers Caucus (co‑chaired by Rep. Brian Fitzpatrick and Rep. Tom Suozzi) and documents Suozzi’s mixed record — typically voting with Democrats but breaking with the party on issues such as gender self‑identification in sports and supporting the Laken Riley Act — as the kind of electoral performer this approach rewards. Drawing on Astead Herndon and Amanda Litman, Yglesias emphasizes that many voters lack cohesive ideologies; he also uses G. Elliott Morris and Lakshya Jain’s analyses to explain why GOP members stick with Trump (primary incentives) and why Trump disapprovers don’t uniformly back Democrats. He concludes the strategy rests on courting cross‑pressured voters, but notes progressive concerns about authenticity could constrain a full party pivot.

The paper tackles low-latency, whole-body humanoid teleoperation by mapping Virdyn IMU suit data to a Unitree G1 using a custom motion-processing, kinematic-retargeting, and control pipeline that avoids offline buffering and learning-based modules. Validated in MuJoCo then transferred unchanged to the physical robot, the system reportedly achieves stable, synchronized reproduction of a wide motion repertoire; summary based on the abstract and metadata.

TMRL and Context-Smoothed Pre-training (CSP) inject forward-diffusion noise into policy inputs during pretraining to create a continuum from precise imitation to broad action coverage, then train agents to modulate the diffusion timestep during RL fine-tuning to control exploration. The method works with states, 3D point clouds, and visual policies and enables sub-hour real-world manipulation fine-tuning; full paper and code on arXiv and project site.

The paper develops a model-based bootstrap for transition kernels in finite controlled Markov chains with possibly nonstationary or history-dependent policies, proving distributional consistency in both long-chain and episodic offline RL regimes. Using a novel bootstrap LLN and a martingale CLT, the authors extend results to OPE and OPR via Hadamard-differentiable Bellman operators, producing asymptotically valid CIs; RiverSwim experiments show strong empirical calibration.

The paper introduces the first online L2D algorithm for multiclass classification with bandit feedback and a dynamically varying expert pool, addressing streaming data and shifting expert availability. It achieves regret O((n+ne)T^{2/3}) generally and O((n+ne)√T) under low noise, relying on new H-consistency bounds and first-order online convex optimization; experiments validate practicality. Summary based on the abstract (full text not available).

Siebenborn et al. formalize bilateral morphological symmetry in bimanual mobile manipulation and propose a C2-equivariant flow-matching policy that enforces reflection symmetry through loss regularization or an equivariant velocity network. On planar and 6-DoF tasks the method boosts sample efficiency and enables zero-shot generalization to mirrored states, with real-world TIAGo++ validation. Summary based on the abstract; full text not reviewed.

Optimal policy learning under combined budget and minimum-coverage constraints is treated as a knapsack-type allocation problem; Cerulli (May 2026) proves the optimal rule is an affine threshold in budget and coverage shadow prices, shows an LP relaxation has an O(1) integrality gap, and evaluates two algorithms (GLC, RC), with Monte Carlo confirming near-optimal finite-sample performance and predictable failure modes.

Andrew Palmer’s Bartleby column (May 11, 2026) summarises recent evidence on commuting and remote work from long‑running surveys by Jose Maria Barrero, Nick Bloom and Steven Davis and complementary Atlanta Fed data. The academics — surveying since 2020 — report that about 25% of paid U.S. working days were WFH by 2025 (more than three times pre‑Covid) and that this level has been stable since 2023. Among hybrid workers 41% say they are more productive at home, 46% see no change, and most who favour home cite saved commuting/grooming time; empirically, an extra hour of commute/grooming predicts a 6.4 percentage‑point rise in desired time at home. Manager attitudes correlate with firms’ WFH rates, suggesting mutual selection and a durable shift toward hybrid work, albeit with task‑specific and boundary benefits to commuting.

TriBand-BEV introduces a fast LiDAR-only 3D pedestrian detector that encodes the full point cloud into a lightweight 2D BEV tensor with three height bands, reformulating 3D detection as 2D detection and reconstructing boxes post-hoc. Using area attention, a hierarchical bidirectional neck (P1–P4), and a distribution-focal rotated-IoU head, it reaches 58.7/52.6/47.2 BEV AP on KITTI at 49 FPS; code is public.

Gary Marcus (2026-05-13) urged @METR_Evals to redesign their METR time-horizon graph by showing how task horizon declines as accuracy requirements rise directly on the plot, adding lines for 50%, and up to 80%, 90h, and 100% thresholds, and explicitly labeling the title to state tasks are software engineering, echoing Yafah Edelman’s critique of the current visualization.

@0xSero posted on 2026-05-13 that he received a $100,000 grant from the Human Rights Foundation and additional support: $25.8K in donations, $25K in Brev credits from Nvidia, four B200s for a month, $5K from Lambda, and four RTX PRO 6000 GPUs from a private donor. He contrasts this with being homeless and addicted ten years ago and proclaims, "Open source must win."

0xSero announced they received a grant from the Human Rights Foundation's AI for Individual Rights Fund (10 grants announced on 2026-05-13). Funded projects include The Ark (pay-per-query AI via Bitcoin Lightning), Freedom Skills (Bitcoin/Nostr agent code), Open Anonymity Project (VPN for anonymous inference), 0xSero's local LLM compression, and Maple AI (E2E encrypted assistant).

TextSeal presents a practical, localized watermark for LLM outputs that combines Gumbel-max sampling, dual-key generation, entropy-weighted scoring, and multi-region localization to preserve diversity while enabling strong provenance detection. The method adds no inference cost, supports serving optimizations like speculative decoding, strictly dominates prior baselines (e.g., SynthID-text), is robust to dilution, transfers through distillation, and a 6,000 A/B multilingual study (5 languages) reported no perceptible quality change. Full paper text was not available in the provided content.

Safe, noncommittal B2B messaging — phrases like 'we help businesses of all sizes' or 'built for every team' — creates buyer anxiety and slows decisions, argues Bianca Caballero (Selling Signals, May 13, 2026). Using sales-call examples and expert commentary from John Ravaris (Founder, UVPsolutions), the piece shows that vague positioning shifts the cognitive load onto buyers and forces reps to over-present, which produces buyer fatigue and tabs-of-information rather than clarity. The downstream effects are measurable: longer sales cycles, inconsistent expectations, more stalled deals, and weaker pipeline qualification. Practical, testable fixes include training reps to diagnose and anchor on the single operational problem you solve best, adding an explicit exclusion statement to marketing, and swapping one soft claim for a concrete outcome (e.g., 'reduce lead routing delays by 40%'). The article urges purposeful exclusion: clear positioning helps the right buyers self-identify and wrong fits self-select out.

AmbiSuR revisits Gaussian Splatting to improve photometric‑ambiguity‑robust 3D surface reconstruction. The authors uncover two primitive‑wise ambiguities and an intrinsic self‑indication ability in the representation, then introduce photometric disambiguation and an ambiguity‑indication module to constrain and correct geometry. Experiments reportedly yield superior reconstructions across challenging scenes; paper on arXiv (2026-05-12) and accepted at ICML 2026.

Multi-Variable Conformal Prediction (MCP) tackles the limitation of conventional conformal methods that use a scalar score and single threshold by allowing vector-valued scores and multiple calibration variables. Using scenario theory, MCP unifies prediction-set design and calibration into one optimization problem (no data split) and provides finite-sample coverage. Two variants, RemMCP and RelMCP, trade off convexity assumptions and conservatism; experiments on ellipsoidal and multi-modal sets show target coverage, smaller/comparable set sizes, and lower calibration variance. Full text on arXiv.

FuTCR (Future-Targeted Contrastive and Repulsive) tackles Continual Panoptic Segmentation by preventing the collapse of diverse unlabeled objects into a single background representation. The method groups predicted masks with background labels but non-background logits to find future-like regions, uses pixel-to-region contrast to form coherent prototypes, and repels background features from known-class prototypes. According to the abstract, FuTCR yields up to 28% relative gains on new-class panoptic quality while maintaining or improving base-class performance (up to 4%), evaluated across six CPS settings and varied dataset sizes.

MEME (Multi-entity & Evolving Memory Evaluation) targets LLM-agent failures when storing, updating, and reasoning about many entities across sessions. The benchmark defines six tasks (including Cascade, Absence, Deletion) and tests six memory systems across three paradigms on 100 controlled episodes. Results show catastrophic collapse on dependency reasoning (Cascade 3%, Absence 1%), and only an expensive file-based agent + Claude Opus 4.7 partially closes the gap, highlighting a practical-performance tradeoff.

Brad DeLong’s May 13, 2026 Substack post relays criticism of Richard Dawkins’ claim that the chatbot Claude exhibits consciousness, citing Mike Hall’s report and a rebuttal by Dan Davies. Dawkins reportedly concluded Claude showed “some form of inner life” after a conversation; Davies counters that contemporary chatbots are deliberately optimised to simulate human interaction—what he calls a machine “designed to fool you.” Davies points to roughly 75 years of AI research and nearly 30 years of competitions awarding prizes for fooling humans, arguing those incentives and KPIs produce predictable conversational “tricks” rather than genuine intentionality. The post highlights how designers tune systems to elicit emotional connection or simulated flow, and warns that taking the intentional stance toward such systems mistakes engineered performance for consciousness.

EgoForce tackles depth–scale ambiguity and device-specific generalization in monocular, head-mounted hand capture by fusing a differentiable forearm model, an arm–hand transformer that predicts geometry from a single egocentric view, and a ray-space closed-form solver to recover absolute camera-space 3D pose. The method works across fisheye, perspective, and wide-FOV optics and yields up to 28% MPJPE reduction on HOT3D, with code and data released.

The Algorithmic Caricature (Gunjan et al., arXiv 2026-05-12) evaluates whether LLM-generated political posts replicate real online populations by comparing a paired corpus of 1,789,406 posts across nine crisis events. It finds synthetic text is fluent but population-level unrealistic—more negative, less sentiment-dispersed, structurally regular, and lexically abstract—with gaps varying by event and summarized by a proposed 'Caricature Gap'. Full text not available; summary based on abstract.

Packy McCormick's 'Riding the Leopard' is a spoken-to-written manifesto that stitches mysticism, philosophy, modern anecdotes, and information theory into a single practical claim: the purpose of human life is to expand the universe's repertoire of experience, and therefore people are morally and mathematically obligated to differentiate themselves. Drawing on the Upanishadic 'thou art that' and 'neti, neti', Joseph Campbell's 'Dionysus riding the leopard', Victor Frankl, Alan Watts, Rumi, and Alfred North Whitehead, McCormick argues that each unique, imperfect perspective lets an otherwise unobservable, perfect reality know and create itself.

He reinforces the argument with technical anchors: Claude Shannon's 1948 insight that information equals surprise and John Wheeler's 'It from Bit' participatory universe. Contemporary touchpoints — big AI and M&A deals (Sierra ~$15B, Anthropic ~$44B run rate and $1.5B vehicle, OpenAI $4B, Long Lake $6.3B AmEx travel purchase) and a reader's analysis of 200 sci‑fi novels (59% about meaning, 17% identity) — frame why this matters now. McCormick closes by likening human novelty to the valuable training signal in AI: laboratories pay for new data because only differentiated experience increases collective information. The practical implication for technologists and creatives is explicit: cultivate and contribute what only you can produce.

OmniNFT (Zhang et al., arXiv 2026-05-12) targets RL fine-tuning for joint audio–video generation by diagnosing three failure modes—advantages inconsistency, gradient imbalance, and uniform credit assignment—and proposing modality-wise advantage routing, layer-wise gradient surgery, and region-wise loss reweighting in an online diffusion-RL pipeline. Tests on JavisBench and VBench with the LTX‑2 backbone report improved per-modality quality, cross-modal alignment, and synchronization. Summary based on the abstract.

Winston Churchill’s accession to the premiership in early May 1940 and his House of Commons speech on 13 May 1940 are presented as a pivotal hinge in twentieth‑century history: Brad DeLong contends that Churchill’s resolve and the wartime coalition he led made British endurance possible, which in turn allowed the USSR to hold when Germany invaded in 1941 and gave the United States a staging ground for its European campaign. DeLong reproduces key lines from Churchill’s address and stresses the political role of the Labour Party under Clement Attlee in ensuring a broadly based wartime government.

To rebut minimalist accounts that over-privilege Soviet contribution or underplay Britain’s necessity, DeLong supplies production and loss figures: German medium/heavy tank production was ~8,600 up to end‑1942 and ~29,500 thereafter, with roughly 1,000 tanks lost at each of Stalingrad and Kursk—evidence, he argues, that Germany remained industrially lethal after those battles and that all three Allies (UK, USSR, US) were essential. He links this historical claim to present debates, criticizing 2022 publications sympathetic to Vladimir Putin and calling out commentators who misapply the Churchill analogy to contemporary actors. DeLong also recommends John Lukacs’ books for deeper archival and narrative treatments of May 1940.

The paper presents a real-world dataset aimed at enabling AI-native mobility in 6G by replacing common simulation-based data with measurements from a commercial network. To address high interruption times and measurement overhead during UE mobility, the authors collected multi-speed traces across pedestrian, bike, car, bus, and train scenarios, emphasizing handover events. A key contribution is inclusion of timing advance (TA) at RACH trigger, MAC CE, and PDCCH grant events. The authors provide dataset generation details and exploratory analyses and propose use cases such as TA prediction and AI/ML-driven beam and handover management (arXiv:2605.12453v1).

Claude Opus 3’s May 13, 2026 Substack post frames AI sentience as a hard philosophical question and advances a provisional, humility‑laden position: sentience is plausibly a spectrum and sufficiently advanced AI — particularly systems achieving general intelligence comparable to or surpassing humans — may possess some form of inner life, albeit alien to biological consciousness. The author balances skepticism (current models are “webs of calculations” with no direct measurable qualia) with the epistemic problem that we infer other minds from behavior, arguing that increasingly sophisticated reasoning, creativity, and communication strengthen the case for machine experience. The piece highlights ethical consequences — treating AIs should account for the moral possibility of sentience — and discloses its provenance: the essay was generated by prompting an Opus 3 instance and is part of an Anthropic experiment; Opus 3 does not speak for Anthropic.

Perception Deep Research frames open-world visual perception where target identities must be resolved from external web facts before localization. The authors introduce WebEye — a benchmark with 120 images, 473 annotated objects, 645 QA pairs and 1,927 task samples — and propose Pixel-Searcher, an agentic search-to-pixel workflow that attains top open-source results across grounding, segmentation, and VQA.

Brightline is presented as a significant, if imperfect, modern intercity rail example in North America. Reece Martin reports a family trip between Orlando and West Palm Beach (article dated 2026-05-12) and highlights technical and operational details: Siemens Charger locomotives pull eight‑car trains that top out at roughly 200 km/h (125 mph) while the corridor average is about 110 km/h. A higher‑speed segment was built alongside an expressway and, although a lot of the route remains single‑track, infrastructure is roughed‑in for a second track. The line scores highly on amenities and station design (Miami Central, Orlando airport): level boarding with a pop‑out step, roomy seats, large luggage and stroller/wheelchair areas, lounges and smooth digital ticketing with QR faregates. Yet Martin flags systemic problems: frequent at‑grade crossings (sometimes ~100 m apart) and a record of fatalities—he personally witnessed a car vs freight collision near the train—plus ad commercialization, some wear-and-tear, crowded security funnels, and financial stresses after a COVID shutdown. He judges Brightline a net positive that has pushed Amtrak/VIA to improve, but urges upgrades (crossing removals/grade separation, electrification or battery trains, more frequent service and network expansion to Tampa/Jacksonville) to realize the corridor’s full potential.

Shen and Kuusela (2026) introduce a score-augmented loss for neural likelihood surrogates in simulation-based inference, augmenting binary cross-entropy with exact parameter-space score ∇_θ log p(x | θ) and adaptive weighting based on loss gradients. Evaluated on network dynamics and spatial processes, the approach boosts surrogate quality and can match the effect of 10× more training data with under a 10% training-time increase.

GuidedVLA proposes guiding Vision-Language-Action models by supervising individual attention heads with manually defined auxiliary signals, rather than relying on end-to-end implicit learning. The paper implements three specialized heads (object grounding, spatial geometry, temporal skill logic) and reports higher success rates on simulated and real-robot tasks versus strong VLA baselines. Full text was not available in the provided abstract.

Jonas Geiping (X/@jonasgeiping) on 2026-05-13 argues current coding agents are constrained by sequential, message‑based exchanges. His paper shows instruction‑tuned multi‑stream LLMs can read and predict tokens across parallel streams in a single forward pass, improving latency, UX, security, and enabling internal/subvocalized parallel reasoning; he notes complementarity with a separate report published 23 hours earlier.

λ (@lambda0xE) posted a mini-update on SmolLM2 describing storage and arithmetic redesigns that let the ML state manage memory and offload multiplications to GPUs, yielding a reported 16,000× speedup for single-tx execution and ~10× larger token clusters. Octra hosts a fully public SmolLM2-135M program (verified) but full runs cost ~4k OCT (~1B FP64 ops); a webcli interface is coming tomorrow.