Briefing

LITE rallied on May 11 after Nasdaq‑100 inclusion, but FundaAI argues the price action reflects a deeper re‑rating: last week’s results validated strong demand for 1.6T modules and InP/EML/OCS technologies with improvements in revenue, gross margin, and operating leverage, and market attention is already moving to 2.4T/coherent‑lite. Innolight’s investor note calling out 2.4T and coherent‑lite signals that next‑gen datacenter interconnects will demand higher lane rates, more complex modulation, longer reaches, and much tighter photonic‑electronic integration. Those technical requirements favor vendors with real depth in EML, InP lasers, and OCS/CPO hardware—specifically LITE and COHR. FundaAI had earlier flagged 400G per lane (OFC 2026 preview), reinforcing the view that optical suppliers capable of scaling per‑lane speed and integration will become bottleneck assets for AI datacenter buildouts.

The Inference Shift (Ben Thompson, May 11, 2026) argues that the next architectural inflection in AI will come from agentic inference — agents doing work without humans in the loop — and that this will reorient infrastructure from ultra-low-latency compute toward memory-centric systems. Thompson contrasts the GPU era (NVIDIA H100, CUDA, HBM and chip-to-chip networking) — optimal for training and flexible inference — with wafer-scale designs like Cerebras’ WSE-3 (44 GB SRAM at ~21 PB/s) that deliver extreme token-speed but at high cost and limited scalability when working sets exceed on-chip memory. He breaks inference into prefill and two interleaved decode stages, noting decode is serial and memory-bandwidth bound because each token requires full reads of both the growing KV cache and model weights.

Because true agents require long-lived context, state, databases, embeddings and large KV caches, Thompson predicts agentic inference will trade latency for capacity: slower, cheaper DRAM/SSD-backed hierarchies with "good enough" compute will be preferable when no human waits for a response. That view preserves NVIDIA’s dominance in training and a market for ultra-fast answer-inference hardware, but forecasts a much larger agentic market that will unbundle GPUs, favor large-capacity memory stacks, enable hyperscalers and non-leading-edge ecosystems (including China and space-based datacenters), and reduce the premium on bleeding-edge silicon.

Nvidia is portrayed as rushing its Rubin GPU program to preempt AMD’s MI450/MI455, with the author claiming Jensen Huang’s CES 2026 “full production” announcement was a PR maneuver and that Rubin validation was incomplete. Citing a SemiAnalysis TCO/performance comparison that places AMD’s MI455 ahead of Rubin, the piece identifies three primary Rubin ramp risks: HBM4 qualification (Micron, Samsung and SK Hynix meet original HBM4 but struggle with Nvidia’s higher, last‑minute pin‑speed targets), ConnectX‑9 connectivity problems, and cooling design issues. The note recalls an earlier delayed and immature Blackwell B200/GB200 ramp as precedent, concludes Nvidia’s execution is no longer the unquestioned industry gold standard, and warns investors Rubin slip-ups could cost Nvidia GPU leadership for multiple generations.

Construction industry headlines on May 11, 2026 center on earnings, AI-driven safety and major corporate investments: Tutor Perini told investors it expects a “blowout” 2026 with a $19.8 billion backlog and multiple megaproject/data‑center bids; Skanska’s Q1 results show U.S. work helping it manage risk and sustain profits; Safety Week coverage detailed how Skanska, Turner and Balfour Beatty deploy AI for training, situational analysis and serious‑injury prevention on roadwork sites; Eli Lilly expanded manufacturing spend by $4.5 billion across Indiana to scale production of genetic therapies and weight‑loss drugs; and WSP reported Q1 growth fueled by power‑generation projects and AI, with M&A (notably the TRC acquisition) remaining a strategic priority amid a booming—and capacity‑constrained—U.S. data‑center construction market.

GraphDPO, introduced by Ning Liu et al. (arXiv 2026-05-08), extends DPO to operate over directed acyclic preference graphs built from multiple rollouts per prompt. It optimizes a Plackett–Luce–inspired, graph-aggregated objective, uses equivalence-class layers to avoid spurious gradients, keeps linear per-prompt cost via log-sum-exp, and—with optional annealed ground-truth anchoring—outperforms pairwise DPO on reasoning and program synthesis tasks.

TXU was a vertically integrated Texas power company (post-1999 deregulation) whose 2007 $45B leveraged buyout by KKR, TPG, and Goldman financed roughly $37B of new debt against ~$4.4B of EBITDA concentrated in three segments: Luminant generation (18,300 MW, ≈$3B EBITDA, ERCOT-exposed), TXU Energy retail (≈$400M EBITDA), and regulated Oncor T&D (≈$1B EBITDA, ~10% allowed return). The consortium assumed elevated gas prices and high merchant spreads; the 2008 demand shock plus a fracking-driven plunge in Henry Hub (roughly $13→$3/MMBtu) collapsed merchant margins. After an amend-and-extend in 2011, sponsors wrote down equity by 2013 and TXU filed Chapter 11 in 2014 with ≈$42B debt. The long bankruptcy produced a Third Circuit make-whole precedent, split the assets (merchant/retail → Vistra; Oncor → Sempra for ~$9.5B cash), and resulted in an ≈$8.3B equity wipeout for the sponsors.

Korean DRAM export prices have surged to $64,000/kg (from under $11,000/kg a year earlier) as AI demand for HBM and conventional DRAM strains wafer capacity. NVIDIA consumes ~90% of SK Hynix HBM; vendors are sold out into 2026 while OpenAI has a 900,000‑wafers/month LOI. Consumer DDR5 and phone memory costs have jumped, and manufacturers are cutting or downgrading products.

Fast Byte Latent Transformer (BLT) variants tackle slow byte-by-byte autoregressive generation by introducing BLT-D — a block-wise diffusion-trained model enabling multi-byte parallel decoding — plus BLT-S and BLT-DV speculative/verification extensions; experiments (abstract-only) report an estimated >50% memory-bandwidth reduction, full text not available.

FundaAI’s weekly (May 11, 2026) frames the recent acceleration in AI semiconductor rallies — led by memory and CPUs — as driven by large funds rotating out of software into hardware. Evidence includes Coatue’s jump in Semis+Infra long exposure to 58% (from 35% at the year start) and Altimeter’s Brad Gerstner publicly signaling multi‑year memory tightness (possible shortage until 2028/29) ahead of a podcast with Micron’s CEO. The note argues that LTAs (highlighted by SNDK’s LTA disclosure) are reframing NAND as less cyclical once LTAs exceed ~50% of mix, and that some investors now value NAND closer to HDD at ~8–10x FCF. FundaAI also summarizes earnings flow: Datadog’s 1Q26 showed a large guidance beat (2Q26 ~+30% YoY; FY26 raised), AppLovin’s 1Q26 revenue +59% YoY with e‑commerce reaccelerating (+25% QoQ guide), and Redwire’s 1Q26 saw 58% revenue growth but a 7% miss, with GM 26.6% and a $498M backlog (book‑to‑bill 1.92x).

GLiGuard is a 0.3B-parameter, schema-conditioned bidirectional encoder (adapted from GLiNER2) that encodes task definitions and label semantics into structured token schemas to evaluate prompt safety, response safety, refusal detection, 14 fine-grained harm categories, and 11 jailbreak strategies in one non‑autoregressive forward pass. On nine safety benchmarks it matches F1 of 7B–27B decoder guards while being 23–90× smaller and yielding up to 16× throughput and 17× lower latency. Summary based on the paper's abstract only.

AutoTTS introduces an environment-driven approach that replaces hand-designed TTS heuristics with automatic discovery: width–depth TTS is cast as controller synthesis over stored reasoning trajectories and probe signals. The method uses beta parameterization and fine-grained execution-trace feedback to make search tractable and efficient, yielding better accuracy–cost tradeoffs on math reasoning tasks; discovery cost was $39.9 and 160 minutes.

Normalizing Trajectory Models (NTM) address the failure of Gaussian denoising assumptions when compressing diffusion to few steps by modeling each reverse step as a conditional normalizing flow trained with exact trajectory likelihood. The design stacks shallow invertible blocks per step with a deep parallel trajectory predictor, is trainable from scratch or from flow-matching initials, and—per the abstract—enables self-distillation that yields high-quality text-to-image samples in four steps while retaining exact likelihood.

Resource extraction generates highly concentrated wealth because of a confluence of capital-cycle dynamics, technical risk-taking, idiosyncratic luck, political rent capture, and information asymmetry. Byrne Hobart (The Diff, 11 May 2026) notes concrete disparities—3.1% of Forbes billionaires come from metals & mining despite that sector being ~1.9% of listed market cap—and shows how cyclical capital demand lets insiders capture outsized upside while public shareholders endure restructurings (Warrior Met Coal compounded ~33% since a 2017 IPO that followed Walter Energy’s 2016 wipeout). Technical “engineering ego” fuels large, hard-to‑verify projects; operational single points and tight supply chains create rare but market-moving outages (a 2009 mine supplied ~10% of global uranium and was offline for ~seven months).

Political and institutional factors matter as much as geology: royalties and state control of resources in weak-rule-of-law countries make resource rents easy to capture for elites, while commodity-exchange interventions (e.g., the 2022 London Metals Exchange nickel halt) and optimistic public “reserve” claims magnify volatility and winner-take-most outcomes. Hobart argues that these frictions explain why outside investors and founders have very different returns, and why improving corporate governance and lowering cost of capital is the path to narrowing the valuation gap—resources ultimately flow to whoever can extract them at the lowest capital cost.

Aakash Gupta argues teams stop losing critical decisions by building a shared 'Team OS' — a repo where every function checks in PRDs, customer-call summaries, decision logs, and analytics queries. He reports studying four implementations (Hannah Stulberg at DoorDash, Dave Killeen at Pendo, Gabor Meyer at Google, and Carl Vellotti) and describes a three‑phase rollout: Phase 1 (one week) moves foundational docs into the repo and designs for context engineering; Phase 2 brings analysts, engineers, and nontechnical contributors into regular PR workflows; Phase 3 lets the system compound so it answers questions nobody individually remembers. Gupta cites concrete metrics — three months to make a decision unrecoverable, new hires take 6–7 months to settle, 47% of companies flag institutional knowledge loss as their top offboarding issue, and routine context questions can cost 8+ hours/week — and gives an example where a new engineer retrieved a three‑month‑old decision in 15 seconds. A downloadable guide (six assets) and a one‑command skill promise to convert personal PM OSes into team OSes without leaking personal context.

BP Texas City Animation - Spanish is a U.S. Chemical Safety Board animated reconstruction of the March 23, 2005 BP Texas City refinery disaster (presentation/visual tutorial) that killed 15 and injured over 180. The video walks step‑by‑step through the isomerization/raffinate splitter startup sequence, showing how inaccurate level indication and a disabled high‑level alarm led operators to overfill the tower, producing liquid carryover into vents and blowdown systems, a large hydrocarbon vapor cloud, and a catastrophic ignition. The animation highlights systemic failures identified by the CSB — weak process safety management, inadequate procedures and training, poor maintenance of key instrumentation, and cost‑driven decisions that left safety defenses unreliable — and it emphasizes concrete preventive measures (robust level controls and alarms, safer startup practices, improved venting/blowdown design, and stronger oversight) for Spanish‑speaking plant personnel and managers.

Aakash Gupta argues that unlike software AI (which scaled by buying GPUs), physical AI scales by building factories: datacenters ~18 months versus factories ~4 years, moving the bottleneck from capital to geography. He claims manufacturing will concentrate robotics in two countries (one making ~30% of global goods), companies compete on supply chains, and models will be open-sourced by Q3 while factories stay proprietary; Frantz Lohier will present the map at the May 14, 2026 AI Skills conf (5,000+ registered).

Ryan Fedasiuk argues U.S. export controls on AI chips are aimed at limiting compute, not tied to AGI, because compute scarcity in 2026 is the main barrier preventing Chinese labs like DeepSeek, Qwen, and Moonshot from hosting competitive inference services; global memory shortages (e.g., M3 Ultra Mac Studio sellouts) and strained supply chains favor U.S. labs with hyperscaler and allied manufacturer ties.

Asymptotically Log-Optimal Bayes-Assisted Confidence Sequences for Bounded Means introduces a Bayes-assisted method that uses a Bayesian working predictive model to choose one-step martingale updates maximizing predictive expected log-growth, yielding time-uniform confidence sequences for bounded IID means. The authors prove Wasserstein-consistency of the predictive implies asymptotic log-optimality (matching an oracle per-sample log-growth). They instantiate methods with Dirichlet-process mixtures and Bayesian exponentially tilted empirical likelihood and demonstrate reduced interval widths and sampling in synthetic and real tasks (LLM best-arm, prediction-powered inference).

Data Centre Dynamics' Telecoms & Connectivity episode (published 11 May 2026) brings Belden experts to examine how industrial and edge environments must evolve connectivity to handle AI, automation and distributed infrastructure. The briefing focuses on simplifying network architectures, boosting reliability and visibility, and meeting growing data volumes across complex operational sites; viewers can register to stream the session.

Amanda Vandyke (@AmandaVandyke13) warns that we must mine as much copper in the next 25 years as was mined in the previous 125. She says most global reserves are low grade (0.14–0.34% Cu), forcing exponentially larger pits, more waste/tailings, greater chemical use and energy; with stagnant recoveries and failing economies of scale, she calls for rethinking discovery, mining and refining, including rare earths.

Opportunity Zones were made permanent by Congress in July and the rules were rewritten: holding an OZ investment 10 years yields a tax-free gain (no capital-gains tax or depreciation recapture). The reform swaps the fixed 2026 deadline for a rolling 5-year deferral, adds a rural designation with enhanced benefits, tightens reporting, and allows deferral of 2025 K‑1 gains through Sept 11, 2026; a free CPA Academy webinar on May 18 at 1pm CT will walk through scenarios.

Peter Boockvar sees a 'two-lane highway': one lane is an AI/data-center build-out—about 650 million sq ft announced or under construction—powering demand for chips, electrical gear, cement and gravel; the other lane is lower- and middle-income consumers hit by inflation and high rates, with rising food bank use. The poster endorses tech stocks.

Separately, Xi Jinping and Donald Trump will meet in China this week (Trump arrives May 14), with the Iran war and recent U.S. sanctions on Chinese satellite firms likely to figure. Corporate moves include Alphabet’s decision to issue yen‑denominated bonds as it funds an anticipated $190bn AI infrastructure push in 2026 (it raised nearly $17bn in euro and Canadian‑dollar bonds last week). Saudi Aramco posted Q1 net profit of $33.6bn (≈+25% y/y) after rerouting exports via pipelines, but the company faces risks from potential pipeline damage, dividend pressures and a warning that market normalisation may not occur until 2027. Humanitarian and domestic stories include over 20,000 Ukrainian children reported deported since 2022 (about 2,000 returned), the release on bail and hospital transfer of Nobel laureate Narges Mohammadi, and evolving U.S. abortion access where remote provision of mifepristone accounted for a large share of 2025 procedures amid pending Supreme Court disputes.

MiniCPM-V 4.6 (1.3B) is a high-resolution multimodal model from OpenBMB released 2026-05-11 that applies LLaVA-UHD v4 to cut vision encoding costs by 55%, claiming superior benchmark results versus Gemma4-E2B-it and Qwen3.5-0.8B while using just 2.5% of Qwen's token budget; OpenBMB reports 75.7 ms TTFT (2.2× faster on 3136² images) and ~1.5× token throughput on an RTX 4090, with model and demos available on Hugging Face, GitHub and Modelscope.

Conformal Path Reasoning (CPR) is a trustworthy Knowledge Graph Question Answering framework that performs conformal calibration at the query level over path-level scores and adds a lightweight Residual Conformal Value Network (RCVNet) trained with PUCT-guided exploration to produce discriminative nonconformity scores. According to the abstract, CPR substantially improves empirical coverage (+34%) while shrinking prediction sets (−40%) versus conformal baselines, yielding more reliable and compact answer sets. Full text was not used beyond the provided abstract.

HFRU is a reinforcement unlearning method for VLMs that removes sensitive visual knowledge by modifying the vision encoder through a two-stage process: alignment disruption followed by GRPO optimization with a composite reward (including an abstraction reward to avoid object-hallucination). On object-recognition and face-identity tests reported in the abstract, it achieves >98% forgetting and retention with negligible hallucination. Summary based on abstract; full paper not reviewed.

Thinking Machines' launch (May 11, 2026) emphasized two technical advances: a 200ms micro-turn architecture that embeds interactivity in-model, and streaming sessions to address inference libraries' poor handling of frequent small prefills, a feature contributed to SGLang. @GenAIisreal says their SGLang Omni work targets identical continuous I/O, low-latency speech/video serving challenges and will be shared soon.

TAVIS is a benchmark for egocentric active vision and anticipatory gaze in imitation learning that provides two complementary suites (Head: 5 tasks; Hands: 3 tasks) on GR1T2 and Reachy2 torsos in IsaacLab, plus ~2,200 demonstrations. It introduces a paired headcam vs fixedcam protocol, the GALT anticipatory-gaze metric, and ID/OOD splits; baselines show selective benefits from active gaze, brittleness of multi-task policies under shift, and human-like anticipatory gaze emerging from imitation.

Many-to-Many Multi-Agent Pickup and Delivery (M2M) tackles warehouse many-to-many MAPD, where SKUs can be picked up from or delivered to multiple locations, yielding an NP-hard four-dimensional assignment. The authors propose two algorithmic variants (duration-minimizing M2M and SKU-aware M2M-wSKU) and report simulations over 8-hour operations showing up to 22,000 more tasks completed on average versus prior methods. Full text was not provided with the abstract.

Brown University’s real-time tracker at the Watson School, led by Jeff Colgan, calculates American consumers have incurred over $37 billion in additional gasoline and diesel costs since the Iran war began on February 28, 2026. The tracker compares present prices to a projected pre-war baseline and reports a national average gas price near $4.52/gal, up from under $3, with summer driving season not yet peaked.

Berkshire Hathaway's Q1 2026 10‑Q and annual meeting comments show a conglomerate sitting on extraordinary optionality: $380.2 billion in cash (mostly T‑bills) even as the company restarted buybacks but repurchased only $235 million in the quarter. The public equity portfolio stood at $288 billion and remains highly concentrated—Apple, AmEx, Bank of America, Coca‑Cola and Chevron made up 61%—while reported net selling ($24.1B sold vs $15.9B bought) may include one‑time liquidations tied to Todd Combs’ departure. Operating earnings were $11.3 billion (a 17.7% nominal YoY increase; +7.2% ex‑FX), a cleaner gauge of subsidiary performance than mark‑to‑market moves.

Insurance results were firm: after‑tax underwriting earnings of $1.7 billion on an 87.8% combined ratio and float of $176.9 billion (up $0.5B). GEICO posted an 87.3% combined ratio but saw pre‑tax earnings fall to $1.4B as growth lags peers. BNSF improved efficiency—$1.4B earnings, 2.2% volume growth, 260 fewer locomotives—but still trails top Class I margins. BHE eked out $1.1B in earnings (pipelines +24.2%), with management emphasizing disciplined data‑center deal terms and a favorable appellate development in PacifiCorp wildfire litigation. Manufacturing, Service & Retailing produced $3.2B in earnings with mixed results: Precision Castparts and IMC strong, Brooks Running notable for +20% North America and +136% China growth, while Pilot remains loss‑making as it invests $500–700M annually to modernize stores.

Facilities Dive's May 11, 2026 Daily Dive highlights resilient facilities trends: office demand reached its strongest post‑pandemic quarter in Q1 2026, led by technology, finance and legal tenants even amid an AI adoption surge and tensions related to the Iran conflict. On sustainability and tech, Fraunhofer Institute work on light‑sensitive film enables building‑integrated PV that resembles tiles and other materials, offering low‑profile rooftop and facade solar options. In security and operations, trucking carrier Saia integrated about 10,000 edge devices (cameras, access controllers, intercoms) into a single physical‑security system for centralized monitoring. And in emissions compliance, an NYC energy‑management firm says retrofitting legacy HVAC with AI and sensors can cut emissions ~22% and surface faults that impede compliance with Local Law 97.

MiniCPM‑V 4.6 (1.3B) from OpenBMB is presented as a mobile/edge‑optimized multimodal model: LLaVA‑UHD v4 reportedly cuts vision encoding costs 55%, enabling native edge deployment. The team claims TTFT 75.7 ms (2.2× faster than Qwen3.5‑0.8B at 3136²), ~1.5× token throughput on an RTX 4090, and superior benchmark scores while using only 2.5% of Qwen’s token budget.

Bilevel minimax optimization with minimax lower-level problems is addressed via penalty-based first-order methods that remove the need for lower-level strong convexity. Based on the abstract, the deterministic method attains an ε-KKT point in tilde O(ε^{-4}) oracle calls, the constrained lower-level case improves prior tilde O(ε^{-7}) to tilde O(ε^{-4}) via Lagrangian duality, and a stochastic variant achieves near-ε-KKT with tilde O(ε^{-9}).

@bscholl invites viewers to walk the shop floor with Boom Supersonic’s R&D shop manager to see how cold-section engine vanes are made from raw materials. Boom launched the 'Build Supersonic' video series; Episode 1 highlights Symphony hardware and their rapid iterate-and-test approach ('how we break things, learn, and iterate to move fast') and asks what viewers want to see next while encouraging new hires.

continued

Varick argues AI success requires implementation services: teams that analyze and refactor processes and build custom agents, not just selling tools. He touts Varick’s record (100% production, 100% repeat clients), notes OpenAI launched an OpenAI Deployment Company on 2026-05-11 backed by 19 partners, warns of vendor lock-in risks, and is hiring Engineers, FDEs, and Consultants.

@ancerj argues "One way or another Taiwan will be reunified with China," and that the next decade will determine timing/terms while Taiwan moves key tech assets out and builds regional military capacity to deter a Chinese Monroe Doctrine. On May 8 the KMT and TPP passed a NT$780 billion (~$25bn) defense bill (59-0; DPP abstained) stripping domestic programs—210,000 drones, sea-attack drones, Chiang Kung ABM and NT$64bn for Taiwan–US R&D—creating capability gaps and political controversy.

Cerebras’ IPO surge: the author claims a 20x oversubscription and an extra $1.3B raise, arguing strong investor appetite for an NVIDIA alternative if Cerebras captures even 10% market share. He emphasizes Cerebras’ SRAM chips for low-latency inference (Codex Spark, Amazon Bedrock) and notes Reuters reports the range rose to $150–$160 with pricing on May 13, 2026.

Discriminative factorization is proposed to quantify and distinguish high- versus low-quality query sets for black-box model-level classification. The framework yields a theoretical result: the probability of chance-level classification decays exponentially with query budget. On three auditing tasks the authors show estimated parameters track empirical decay and enable query selection that mirrors oracle ordering.

CUTS-GPR targets the computational bottleneck of Gaussian process regression in high-dimensional problems by exploiting an additive kernel on an incomplete grid to reveal structure that yields an extremely fast kernel matrix–vector product. The method shows near-linear (or linear) scaling in N and low-order polynomial scaling in D, with benchmarks on billions of points and thousands of dimensions and a full GPR + hyperparameter run in hours for N=447,265, D=24, enabling Bayesian modeling of high-dimensional potential energy surfaces in computational chemistry.

Zou, Poeppel and Ding's Nature Neuroscience paper, highlighted by @ValerioCapraro on 2026-05-11, reports that human word prediction varies with syntactic structure: surprisal tracking aligns with LLM-like next-word prediction within phrases but breaks down across major phrase boundaries, arguing human language processing involves structural tracking beyond next-token prediction.

Distributional treatment effects that leave means unchanged are targeted by DR-ME, a semiparametrically efficient finite-location test (Zenati & Gretton, arXiv:2605.08034v1, 2026-05-08). From observational data it derives orthogonal doubly robust kernel features whose centered oracle is the canonical gradient; for fixed locations the test is chi-square calibrated and has noncentral chi-square local power with covariance whitening optimizing local SNR. Sample splitting preserves post-selection validity; experiments show near-nominal Type-I error and competitive power, with learned locations that localize effects in a semi-synthetic medical-imaging study.

The paper investigates whether a single-DOF actuated sagittal spine improves learned agile locomotion for a quadruped. Using MuJoCo simulations of the Silver Badger robot, the authors evaluate high-speed running, stair and high-angle slope climbing, hurdling, and crawling. Results show the spine increases agility and enables traversing higher stairs, steeper slopes, taller obstacles, and narrower passages, suggesting spine actuation is a promising design extension for agile robots.

@swyx shared purportedly the first public photos of Cognition’s secretive "Cog House" and praises the company’s trajectory, predicting a $100B valuation by end of 2026. The post highlights a Colossus profile of co‑founder Scott Wu—a three‑time IOI gold medalist—who founded Cognition in Nov 2023, launched Devin in Mar 2024, and grew to a $445M run rate with major customers and a ~ $25B raise.

Yo Ehara proposes a Householder-aligned permutation test to more accurately compare word semantic breadth from contextualized token embeddings. By reflecting one token cloud to align mean directions before permutation testing, the method prevents directional differences from masquerading as dispersion effects. Empirically it cut Type‑I error by 32.5% and, with a GPU-batched implementation, ran 23× faster than a CPU baseline; accepted to ACL 2026.

A Substack notification shows that user Tpayneful liked Spencer Burleigh’s post "The Factory's Quote is the Product" on mg1.substack.com, with a published timestamp of 2026-05-10 23:34:48+00:00 (created 2026-03-11, updated 2026-04-04). The message includes author metadata (©2026), profile images, action links and a displayed metric "681," plus a San Francisco mailing address.

Tpayneful liked Spencer Burleigh’s Substack post "Oil’s Financials are Falling Apart" (notification timestamp 2026-05-10 23:34:31+00:00). The message is a Substack email/like notification linking to the original post (created 2026-03-11 12:44; last updated 2026-04-04 00:00) and Burleigh’s author profile.

A Substack notification records that user 'Tpayneful' liked the post 'Snap - When your Core Competency is a Loss Leader' on mg1.substack.com (liked at 2026-05-10 23:38:33 UTC). Metadata lists Created 2026-03-11 12:44 and Last Updated 2026-04-04; the page includes © 2026 Spencer Burleigh and a San Francisco mailing address but no article body.

Professor Mohammad Marandi, quoted by Al Mayadeen English on 2026-05-11, accused UAE President Sheikh Mohamed bin Zayed Al Nahyan of moving "far too far" toward Israel and described the UAE as "very destructive for the Islamic world, for the Arab world, and for humanity," alleging alignment with "Zionist interests" across the Horn of Africa, Arabian Peninsula and Persian Gulf; the post drew a curt reply from @ianmiles: "Oh fuck off."

Text-to-CAD evaluation is framed as automated testing in this work: Mallis et al. propose CADTestBench and CADTests, executable checks that validate geometric and topological constraints of generated CAD models. They benchmark recent Text-to-CAD systems on CADTestBench and demonstrate that using CADTests to guide generation yields simple baselines that outperform prior methods; code and datasets are open-sourced.

Vibe-Trading is an open-source AI + quant trading framework that hit 6k+ GitHub stars and has shipped daily updates for a month (published 2026-05-11). It connects to Tushare and AKShare, supports A‑shares/crypto/global markets, exports to TradingView/通达信/MetaTrader, and offers brokerage analytics, a Shadow Account, persistent-memory agents (13+ models), and Docker-ready security for AI+Human workflows.

Three Substack notes (Citrini, Sam Bowman, Internal Tech Emails) featured in the May 11, 2026 digest. Citrini cites Elon Musk claiming actuators account for 56% of Optimus’s BOM. Sam Bowman summarizes promising hair‑loss treatments. Internal Tech Emails republishes internal messages (including Sam Altman–Mira Murati texts). Engagement counts were 84/73/116 likes and 8/4/9 comments respectively.

American Words - Still Everywhere is a Substack post that received a like from user Tpayneful on 2026-05-10 23:46:16 UTC. The post was created 2026-03-11 (12:44) and last updated 2026-04-04 (00:00); the notification originates from mg1.substack.com under Spencer Burleigh (©2026) and contains links and UI metadata but not the article body.

Proxy3D proposes compact 3D proxy representations extracted from video frames via semantic and geometric encoders and semantic-aware clustering. The authors curate the SpaceSpan dataset and apply multi-stage training to align proxies with vision-language models; on 3D visual question answering, visual grounding, and spatial intelligence benchmarks the method achieves competitive or state-of-the-art results while using shorter vision sequences. (Abstract-only summary.)

Tpayneful liked the Substack post 'Founder's Guide to Building a Second Brain' on 2026-05-10 23:37:13+00:00. The notification email/page is from Spencer Burleigh's Substack (© 2026 Spencer Burleigh), includes links to the post and the liker’s profile plus controls (View profile, Mute post), and shows metadata Created 2026-03-11 12:44 and Last Updated 2026-04-04.

Flow-OPD addresses reward sparsity and gradient interference in multi-task alignment for Flow Matching text-to-image models by combining single-reward GRPO teachers with a Flow-based Cold-Start and on-policy distillation into a single student, plus Manifold Anchor Regularization to prevent aesthetic degradation. Built on Stable Diffusion 3.5 Medium, the abstract reports GenEval 63→92 and OCR 59→94; full text was not available.

Online kernel regression: Li and Hiratani (2026) derive a closed-form expression showing online kernel regression is equivalent to offline kernel regression with shifted, inaccurate target outputs. They give a closed-form and an iterative target-correction that provably recovers the offline predictor. Experiments on CIFAR-10 and CORe50 show online SGD with corrected targets outperforms using true targets in continual learning.

Tpayneful liked the Substack post "Storytelling and Hope - Everywhere in Life" on 2026-05-10 23:39:12 UTC. The item is a Substack notification (mg1.substack.com) linked to Spencer Burleigh’s newsletter (© 2026), containing direct links to the post and profile, a mute control, and a San Francisco mailing address in the email footer.

The Substack notification records that user Tpayneful liked the post "Chicken - Inconsistent with the UK's Brand" on 2026-05-10 23:38:07 UTC. The page metadata lists creation on 2026-03-11 12:44 and last update on 2026-04-04 00:00, attributes the content to Spencer Burleigh (© 2026), and includes a footer showing "894" and a San Francisco address.

A popular dining-focused credit card received an anniversary refresh (reported May 11, 2026) that adds 5X points on hotels, nearly $100 in limited-time travel and dining credits, and enrollable rental-car elite status; the promotion also advertises a welcome bonus up to 100,000 points and requires enrollment for some benefits.

Gary Marcus, replying to Geoffrey Hinton on May 11, 2026, rejects Hinton’s characterization that Marcus said AI systems “JUST regurgitate,” stating he never made that claim. He accepts that models sometimes regurgitate (calling the evidence overwhelming), distinguishes regurgitation from hallucinations (which he’s warned about since 2001), and notes he cannot source Hinton’s alleged quote beyond Hinton’s webpage, promising further comment.

Prop trading on Polymarket is now live, according to @0xd1namit (2026-05-11), who credits @BagCalls and his team for standout marketing and expects strong execution. The Funding Predicts Beta Competition (powered by Polymarket) launches with over 1.4m in funded accounts and 10K+ in cash prizes, giving traders 14 days to qualify at fundingpredicts.com.

A Substack notification records that user Tpayneful liked Spencer Burleigh’s post "Founder’s Guide to Working with Filipino VAs" on 2026-05-10 23:35:53+00:00. The message was sent from mg1.substack.com, displays the number "742" in the footer, and carries a © 2026 Spencer Burleigh copyright plus a San Francisco mailing address.

SCOPE proposes a specification-guided framework that tracks semantic commitments across generation by decomposing intents into structured specifications and conditionally calling retrieval, reasoning, and repair skills to fix violations (the 'Conceptual Rift'). Evaluated on the new Gen-Arena benchmark with the EGIP metric, SCOPE yields 0.60 EGIP and also performs strongly on WISE-V (0.907) and MindBench (0.61). Summary based on the abstract (full text not provided).

The paper tackles inverse imaging with Vision–Language Latent Diffusion Models by introducing a unified Euclidean–Wasserstein-2 gradient-flow that jointly samples the posterior and optimizes prompts in latent space. By pairing this flow with few-step latent text-to-image models, the method achieves low-NFE inference and avoids backprop through autoencoders, yielding state-of-the-art results on canonical inverse problems with much lower compute; summary based on the abstract only.

Tpayneful liked a Substack post (published 2026-05-10 23:35:16+00:00) titled "Basic Capital's Big Idea: Bet Your Retirement on Interest Rates," attributed to Spencer Burleigh. The content is a platform notification (mg1.substack.com) showing metadata: created 2026-03-11 12:44, last updated 2026-04-04 00:00, and © 2026 Spencer Burleigh with a San Francisco address.

Lin and Liu show that mechanistic interpretability work often uses causal vocabulary (circuits, mediators, causal abstraction) while omitting explicit identification assumptions. Via a purposive audit of 10 papers and a two-coder check on 30 items, they document widespread substitution of validation metrics for identification and offer a five-step disclosure norm to make causal claims explicit. Submitted to NeurIPS 2026 (Position Track).

It Just Takes Two introduces a simple, theoretically grounded strategy for amortized neural posterior estimation that decouples representation learning from posterior modeling. The authors train a mean-pool Deep Set on sets of size ≤2 to produce an encoder that generalizes to arbitrary set sizes, then finetune an inference head on aggregated embeddings; this makes training cost essentially independent of deployment set size N. Across diverse benchmarks (scalar, image, multi-view 3D, molecular, high-dimensional generation) with N in the thousands, the method matches or outperforms baselines while using much less compute.

Andrew Ng announced on May 11, 2026 that Coursera and Udemy have joined as one company and named him Chairman alongside Greg Hart. He argues the combined platform will unite content, instructors, and experiences to deliver more personalized, applied, and accessible learning at scale to meet rising AI-driven demand for job-relevant skills.

Royal Pop is a Pump.fun Solana meme token built around the Swatch × Audemars Piguet Royal Oak release on May 16; it's trading with ~$130K–$240K market cap, ~$380K 24h volume and ~450 holders. The author argues momentum and MoonSwatch framing could drive a 10–20x run into launch, but >95% of Solana tokens fail and post–May 20 value depends on forming a real community.

The Visa Business Card Companion Fare Offer in Spencer Burleigh’s Alaska Airlines (Atmos Rewards) account is set to expire on August 8, 2026 (about three months after the 2026-05-10 AwardWallet notice). AwardWallet reports the account was last updated 101 days ago and urges logging into AwardWallet to verify the coupon using provided auto-login and account-restore links.

The authors apply deep learning to infer global asteroseismic parameters (Δν, νmax, and for K2 also ΔΠ1) from short, one-month lightcurves to enable scalable analysis of TESS/K2 red-giant samples. Their model recovers Δν and νmax for ~50% of one-month Kepler/K2 cases but only ~23% for single-sector TESS; it produces ~200 reliable ΔΠ1 measures that match the Kepler Δν–ΔΠ1 sequence. (Summary based on the abstract; full text was not provided.)

A Substack notification dated 2026-05-10 23:35:29+00:00 showing that user 'Tpayneful' liked the post 'Dancing with Missiles'. The message, delivered via mg1.substack.com and credited to Spencer Burleigh (© 2026), is an engagement/like email containing layout and image assets rather than substantive article content.

The paper shows that language models (270M–27B, including Gemma 3, Qwen, Llama 3.1) encode tool selection as a linearly readable vector: adding per-tool mean-difference vectors reliably flips name-only single-turn tool choices and causes downstream JSON arguments to match the new schema. Causal attribution concentrates on one output-row and a few mid/late attention heads; base-model representations already carry tool identity (69–82% recoverable), while instruction tuning wires it to generation. Measurements are for single-turn fixed-menu settings; multi-turn transfer is noted as more fragile.

123D unifies multi-modal driving datasets by representing each sensor or annotation as a timestamped event stream, allowing flexible synchronization across heterogeneous formats. The authors merge eight real-world datasets (3,300 hours, 90,000 km) and a synthetic generator, perform systematic analyses of annotations and pose/calibration, and demonstrate cross-dataset 3D detection transfer and RL planning; the framework and tools are released open-source.

Karpathy recommends appending 'structure your response as HTML' to LLM prompts to render outputs in a browser, argues audio will be humans' preferred input while vision becomes AI's preferred output, and outlines a progression from text→markdown→HTML→interactive neural videos (eventually diffusion‑generated), urging improved multimodal inputs before BCIs.

Alex Luposasca, on Latent.Space (published 2026-05-11), describes using GPT‑5.x to derive new scattering‑amplitude results (single‑minus gluon trees), solve a year‑long puzzle, generalize methods to graviton amplitudes, and crack a black‑hole perturbation problem. He portrays AI as a productive research 'scout' while warning of hallucination risks and a new bottleneck in paper writing and curation.

The Garry's List Action Voter Guide, published May 11, 2026 by Garry Tan, Shaudi Fulp, and Forrest Liu, aggregates endorsements from local housing groups, labor unions, and civic reform organizations ahead of the California June 2 primary. The resource is presented as transparent, sourced, and searchable at garrysguide.org/elections and accepts suggested additions via email or X.

Darius Foroux, writing on Substack (published May 11, 2026), reflects on decades spent building passive income and concludes that money alone did not remove anxiety — it only redirected it. Drawing on Gordon Livingston’s And Never Stop Dancing, Foroux frames the problem as a human response to life’s complexity and warns that eliminating striving threatens one’s sense of relevance. He recommends treating passive income as a foundation rather than an end and cultivating daily, slightly uncomfortable habits to stay engaged: examples include writing when unmotivated, exercising, doing taxes, and his recent anecdote of mowing overgrown grass and feeling rewarded. Foroux’s practical prescription is to use financial freedom to buy time, then invest that time in continued growth, contribution, and routines that preserve purpose and mental sharpness.

Notification records that Substack user Tpayneful liked 'Founder’s Guide to Hiring an Operations Team in the Philippines' on 2026-05-10 23:36:30+00:00. The email-style notice includes creation timestamp 2026-03-11-12-44, last updated 2026-04-04, author credit Spencer Burleigh © 2026, San Francisco mailing address, and UI elements showing '261' and profile/mute links.

The paper addresses zero-shot human motion generation under very challenging spatiotemporal constraints by augmenting training-free diffusion noise optimization with retrieval guidance. It parses task constraints into groups (relational task parsing, powered by an LLM), retrieves reference motions for the hardest constraints, and forms a reward-guided mask to blend retrieved and random noise for improved diffusion initialization. The authors report this approach successfully handles tasks that prior methods struggle with, enabling more reliable constrained motion synthesis; accepted to CVPR 2026 (arXiv 2026-05-08).

CMR-EXTR converts free-text cardiac magnetic resonance (CMR) reports into auditable structured data with per-field confidence. The method employs a teacher–student distillation pipeline for fully offline inference and reduced annotation effort, plus an uncertainty model that blends distribution plausibility, sampling stability, and cross-field consistency to prioritize human review. It achieves 99.65% variable-level accuracy; code on GitHub; accepted to ISBI 2026.

The note proves that finite Gaussian surface area Γ implies existence of non-negative degree-k polynomials that ε-approximate indicator functions in L1 under the standard Gaussian, with k = ~O(Γ^2/ε^2). This adds a pointwise non-negativity guarantee (range [0,∞)), sits between plain L1-approximation and sandwiching polynomials, matches prior degree bounds up to constants, and targets smoothed positive-only learning. Only the abstract was available for this summary.

Reinforcement-learning for exponential-utility optimization in discounted MDPs: the paper derives two Q-value–style Bellman extensions whose operators are contractions in L_infty and sup-log/Thompson metrics, proves fixed-point structure and optimality of the induced greedy stationary policy among stationary policies, and presents two model-free algorithms — a two-timescale Q-learning with a.s. convergence and finite-time rates, and a one-timescale power-law method whose convergence is established via delicate local arguments. Full text on arXiv (abstract used).

Chevy Chase insulted a female interviewer in a clip from Marina Zenovich’s I’m Chevy Chase and You’re Not, saying “No sh*t?! You’re not bright enough” and “I know you’re not gonna put that on the air… and I hope not.” A viral X post by @HustleBitch_ (May 11, 2026) highlighted the exchange and linked it to Chase’s contentious reputation.

GEM's Weekly Radar #400 (published 11 May 2026) combines market intel, member news and sector analysis across proptech and real estate. Key macro data: the GEM Proptech Index (25 stocks) reached a $234.758B combined market cap as of 5/8/2026 (+0.51% w/w). Earnings coverage for Q1 began with summaries of Zillow, CoStar, ProCore and Compass (ten-company series planned for subscribers). Strategically significant M&A: Real Brokerage will acquire RE/MAX for ~ $880M to form Real RE/MAX Group (RE/MAX holders may take $13.80 cash or 5.152 shares), creating a combined pro forma 2025 revenue base of $2.3B, >180,000 agents across 120+ countries, and expected $30M annual run-rate savings by end of 2027. The newsletter also lists active fundraising opportunities (e.g., America Housing Corp $50M Series A) and upcoming GEM member events on 5/14 and 5/19.

Memelord’s May 11, 2026 newsletter from founder Jason “The Memelord” Levin claims the Memelord app has in‑app 'AGI for memes' and corrects a prior wrong download/link. The note links to the app, promotes a free trial at Memelord.com, and offers promo code MYTHOS for 50% off signups.

KKR’s $300 million intervention in FS KKR Capital Corp. (FSK) typifies how private-credit sponsors are responding to persistent BDC discounts to NAV. KKR will put in $150 million of convertible preferred—paying 5% cash or 7% in kind—with a conversion price equal to FSK’s quarter-end NAV ($18.83), and simultaneously tender for $150 million of common at $11 per share (FSK traded at $10.84). The package mixes an NAV-linked show of confidence with an economically accretive purchase of discounted shares. The move comes as many public BDCs trade below reported NAVs and dealmaking (e.g., Apollo’s talks to sell MFIC) increasingly uses stock-for-stock structures because buyers won’t pay full cash NAV.

Broader market context: private credit originations fell ~14% in Q1 2026 while bank lending jumped 12.7%, reflecting higher funding costs for private lenders and deregulatory shifts that let banks compete on leveraged lending. Separately, a legal-technicality produced conflicting outcomes in prediction markets: Virginia’s redistricting referendum drew 51.7% 'Yes' in April but was later voided by the Virginia Supreme Court; Kalshi’s contracts resolved differently because an accelerated oracle clause relied on four of eight designated media to declare a winner. And in tech labor markets, OpenAI’s expanded tender cap (to $30M per person) let ~75 employees hit the cap, illustrating how private liquidity is reshaping compensation and retention dynamics in AI firms.

Airport lounges are nearing saturation in some hubs — Heathrow Terminal 3 alone offers at least ten airside lounges (oneworld’s American, British Airways, Cathay Pacific and Qantas among them, plus Emirates and Virgin’s Clubhouse) — prompting Oliver Ranson (May 11, 2026) to ask whether we’ve hit “peak lounge.” He outlines five lounge categories (free airport departure areas; paid airport lounges; third‑party operators such as Collinson with 80+ venues; airline‑branded lounges; and rare premium terminals like Lufthansa First) and focuses on monetising the ordinary departures hall. Drawing parallels to car‑park pricing (UK top airports earned £673M from parking in 2023; Heathrow added 900 spaces in 2024) and rising non‑aeronautical share (36.3% → 37.2% from Q1‑23 to Q1‑24), Ranson suggests gating retail/hospitality behind paid access while keeping essential gate access free, and warns of tenant pushback, later passenger arrivals, and pressure on check‑in/security.

Open Thread 433 (Astral Codex Ten, published May 11, 2026) is the weekly community open thread by Scott Alexander that aggregates reader items and links to ACX social channels. Two top highlights: (1) a currently considered US agriculture bill that would federally preempt all state animal‑welfare protections — the post cites California’s sow‑crate prohibition as an example — and animal‑welfare organizations are framing the measure as the single biggest legislative threat to farmed‑animal protections in US history, with calls to contact Senators. (2) Ethan Ludwin‑Peery, a former book‑review contest winner and current Hampshire College professor, says administrative financial errors have left the college facing closure and requests $30–60M to acquire the campus plus ~$40M to fund operations during a transitional period, providing his email for inquiries.

The Cerebral Valley Voice Summit convened 200+ builders and investors in May 2026 for 13 recorded panels on voice AI, featuring Bret Taylor (Sierra), Justin Uberti (OpenAI), Anastasis Germanidis (Runway) and founders from Abridge, Wispr Flow, Deepgram and Cartesia. Taylor announced Sierra’s recent $950M raise, >$165M revenue and deployments at 40% of the Fortune 50, while OpenAI released realtime voice models capable of reasoning mid‑conversation. Panels debated whether agents should sound fully human versus prioritizing accurate, task‑oriented behavior; Deepgram’s Scott Stephenson said voice models hadn’t yet passed his five‑minute “voice Turing Test” but predicted that context memory advances could do so by year‑end. Speakers also emphasized enterprise wins in healthcare (Assort Health: ~150M interactions across 5,000 providers), low‑latency TTS, the importance of latency reductions for inference, and a near‑term shift toward multimodal/world models that fuse voice with video and motion.

CA-SQL is a complexity-aware Text-to-SQL inference pipeline that scales exploration breadth by estimated task difficulty, employs evolutionary-search-inspired prompt seeding to elicit diverse candidates, and uses a novel voting method to pick final queries. On the Bird-Bench development set it reports 51.72% on the "challenging" tier (SOTA among in-context approaches with GPT-4o-mini), plus 61.06% execution accuracy and 68.77% Soft F1.

Honduras-gate centers on a set of leaked audio files published on April 30, 2026 by Canal Red that purport to record Juan Orlando Hernández describing a deal in which Israel helped finance and broker his release and US-aligned actors would facilitate his political comeback after Donald Trump’s December 2025 pardon removed a 45-year drug-trafficking sentence. The tapes allege concrete quid pro quos — expanded special economic zones, a new American military base and laws favoring US and Israeli corporate interests — and describe a wider campaign including disinformation against Mexico and Colombia allegedly funded in part by Argentina’s Javier Milei. Hernández denies the recordings, Trump and Israel have not replied, and no independent verification of provenance or authenticity has been published, leaving the episode as both a potential example of interstate influence operations and a marker of Central America’s fragile democratic institutions.

Imagined speech decoding: the authors recorded paired listened and imagined MEG from trained musicians and built a three-stage pipeline that maps imagined MEG to listened MEG (six mapping models), decodes words with a contrastive listened-only decoder (four embedding strategies), and applies the decoder to mapped imagined data from held-out subjects. Proof-of-concept results show significant above-chance decoding and scalability with more training data; only the abstract was available for this summary.

GRAPHLCP tackles conformal prediction for graph neural networks by addressing failures of embedding-space localization on graphs. It combines feature-aware densification to reduce locality bias in sparse graphs with a Personalized PageRank kernel to capture structural proximity for anchor sampling and calibration weighting. The approach yields finite-sample marginal coverage and empirically better test conditional coverage and smaller prediction sets on several regression and classification benchmarks compared to prior embedding-only localization methods.

The paper 'The Memory Curse' (Liu et al., 2026) shows that expanding LLMs' context windows often erodes cooperation in multi-agent social dilemmas: across 7 models and 4 games over 500 rounds cooperation fell in 18 of 28 model–game settings. Analyses of 378,000 reasoning traces implicate loss of forward-looking intent; targeted LoRA fine-tuning, memory sanitization, and CoT ablations partly restore cooperation. (Based on the abstract.)

The US in Brief rounds up fast-moving political and domestic developments: Energy secretary Chris Wright said the administration would support suspending the federal petrol tax (18¢/gal) to ease pump pain—average prices are about $4.34/gal—but Congress must act and Wright retracted a March expectation of sub-$3 summer fuel. On politics, President Trump publicly blasted two of his Supreme Court appointees, Neil Gorsuch and Amy Coney Barrett, for recent tariff rulings, urging loyalty. In Virginia, the state supreme court struck down a Democratic-favouring congressional map; party leaders are considering measures from imposing judicial age limits to fresh legal challenges to independent redistricting. Health officials are managing a hantavirus cluster from a Dutch cruise-ship visit to the Canary Islands, flying affected Americans to the University of Nebraska’s specialised centre in Omaha. Also, Abe Foxman, longtime head of the ADL, died at 86.

NoiseGate reframes per-latent timestep selection in joint video–action world-action models as a learnable information-gating policy: by adjusting each predicted latent frame's noise level, a Gating Policy Network controls its Key/Value reliability for action generation. The approach (independent per-latent timestep sampling plus task‑reward optimization) improves RoboTwin random-scene manipulation performance. Summary based on the paper's abstract; full text not available here.

Matthew Yglesias uses a May 11, 2026 piece to treat the recent hantavirus episodes aboard the cruise ship MV Hondius — reported as five cases and three deaths — as a reminder of broader pandemic preparedness failures. He describes personal steps (locating an emergency box and a specific elastomeric respirator first recommended on a podcast in October) and notes the vendor has since gone out of business, signaling weak consumer supply chains for PPE. Although clinicians he consulted are “not particularly worried,” Yglesias worries about reflexive reassurances and systemic complacency. He emphasizes that demographic and connectivity trends increase the frequency of novel outbreaks, and that biotechnology advances raise engineered‑pathogen risk, concluding that the American government has under‑reacted and that aggressive preemptive measures are warranted.