Daily Tech Feed: From the Labs

On April 24, 2026, the White House fired all twenty-four members of the National Science Board by email — the independent governing body of the National Science Foundation, the agency that funded the public internet, the graphical web browser, 3D printing, the Antarctic climate record, and the foundational research pipeline behind modern AI. The firings came twelve days after the NSB objected to the Office of Management and Budget bypassing their statutory approval authority on a $900 million Antarctic research vessel contract. This episode traces the money, the mechanism, and the history of what the NSF built — and what independent scientific oversight was designed to protect.

Between October 2025 and April 2026, cryptographer Daniel Bernstein published a seven-part blog series titled "NSA and IETF" alleging that intelligence agencies are using the IETF standards process to weaken the next generation of internet encryption. The dispute centers on whether the successor to current TLS key exchange should use hybrid post-quantum cryptography — combining classical elliptic curves with the new lattice-based ML-KEM — or ML-KEM alone. The technical stakes are existential: if ML-KEM is eventually broken and the deployed standard is non-hybrid, every session protected by it becomes retroactively decryptable from stored ciphertext. The cost of the safety net is thirty-two bytes. The cost of removing it could be everything.

A controlled experiment pits a neuro-symbolic system against a vision-language-action foundation model on the same robotic manipulation task, same robot, same simulation, same evaluation protocol — and the results are devastating for the foundation model. The paper "The Price Is Not Right", accepted at ICRA 2026 in Vienna, shows that a symbolic planning system trained on one-sixth the data in thirty-four minutes achieves 95% success on robotic Towers of Hanoi where the fine-tuned pi-zero VLA achieves 34% — and on an unseen four-block generalization task, 78% versus zero. The training energy ratio is eighty to one. The inference power ratio is six to one. For structured manipulation tasks, the "just scale it" orthodoxy fails on performance, efficiency, and generalization simultaneously.

Two papers published days apart have reduced the estimated physical qubit count needed to break widely deployed public-key cryptography by roughly two orders of magnitude — from around one million to as few as ten thousand. Together, they compress the timeline for quantum threats to cryptography from "decades away" to "measurable in engineering milestones." The Google paper also introduces the first use of zero-knowledge proofs as a responsible disclosure mechanism for novel cryptanalytic results, proving the existence of optimized attack circuits without publishing them.

Recent advances in foundational models have yielded reasoning systems capable of achieving a gold-medal standard at the International Mathematical Olympiad. We introduce Aletheia, a math research agent that iteratively generates, verifies, and revises solutions end-to-end in natural language, leveraging a novel inference-time scaling law based upon Gemini Deep Think. Aletheia demonstrates several milestones: a research paper generated with no human intervention (Feng2026) calculating eigenweight

TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate closes a gap that has been open since Claude Shannon defined the theoretical floor for lossy compression in 1948. For nearly eighty years, practical vector quantization methods fell exponentially short of what rate-distortion theory says is achievable — either achieving good distortion bounds only through expensive offline training, or running online but paying an exponentially growing quality penalty at higher bit depths. TurboQuant reaches within a constant factor of 2.7× of the information-theoretic optimum, with no training required, at inference time — enabling LLM KV cache compression to 3.5 bits per channel with zero quality degradation and near-zero indexing overhead for nearest neighbor search.

Nvidia committed $26 billion over five years to building open-weight AI models. This episode examines the strategy: open weights as hardware lock-in, the Nemotron Coalition, NemoClaw agent runtime, the Vera Rubin and Feynman hardware roadmaps, and what it means that a chip company is now competing directly with AI labs on model quality.

Every competitive frontier model going forward is sparse — a Mixture-of-Experts architecture where each token activates only a fraction of the total parameters. That decoupling of parameter count from per-token compute sounds like a free lunch. The engineering bill is 88 pages long. NVIDIAs Megatron Core team just published the full receipt: how they solved the memory, communication, and computation constraints that make MoE training at scale fundamentally harder than dense training, and how their open-source framework now runs DeepSeek-V3 and Qwen3 at over 1,200 TFLOPS per GPU on thousands of GPUs.

Every LLM-based text-to-speech system shipping today carries a structural flaw: text tokens and audio frames move at incompatible speeds inside the same model, forcing engineers to choose between reliability, quality, and inference cost. Hume AI's TADA: A Generative Framework for Speech Modeling via Text-Acoustic Dual Alignment eliminates the mismatch entirely — enforcing strict one-to-one synchronization between text tokens and continuous acoustic vectors, producing zero content hallucinations across 1,000+ test samples and running at 5× the throughput of comparable systems.

Rich Sutton published a 1200-word essay in 2019 arguing that 70 years of AI research proved one thing: general methods leveraging computation always beat human-curated knowledge in the long run. Most researchers disagreed. Then the last five years happened. Now Sutton is at Keen Technologies with John Carmack, building something he says the entire current LLM paradigm still gets wrong — a real-time learning agent that never stops training, never freezes its weights, and learns the way an animal learns: by living in the world.

The economics of vulnerability discovery just broke. In twenty minutes, Claude Opus 4.6 found a novel use-after-free memory bug in Firefox — one of the most audited codebases on the internet, backed by millions of CPU hours of continuous fuzzing. That single result is a waypoint on a documented curve: from GPT-4 exploiting 87% of known one-day vulnerabilities with 91 lines of LangChain code in 2024, to Anthropic's red team finding 500+ high-severity zero-days in well-maintained open-source software in early 2026, to a live collaboration with Mozilla that found 22 Firefox vulnerabilities in two weeks. We are in a window: finding is democratized, reliable exploitation still has friction. This episode documents the curve, names what's most at risk, and argues for what defenders must do before the gap closes.

Language models can quote the manual on a bicycle and still miss a broken chain. Beyond Language Modeling: An Exploration of Multimodal Pretraining argues that this is structural, not incidental: text is a lossy compression of reality, and models trained only on it master the description of shadows without seeing the objects casting them. The paper runs controlled, from-scratch pretraining experiments using the Transfusion framework — combining next-token prediction for language with diffusion for vision — across text, image-text pairs, video, and action-conditioned video. The result is four concrete design insights for multimodal architecture, delivered without the confound of inherited language pretraining.

Speculative decoding already beats autoregressive generation — but it still has a sequential bottleneck: verification must finish before drafting restarts. Saguaro (Speculative Speculative Decoding) breaks that dependency by pre-speculating for likely verification outcomes in parallel. Cache hit: return immediately. Cache miss: fall back cleanly. Up to 2x faster than optimized SD, 5x over autoregressive. Lossless by design.

On the night Alibaba shipped Qwen3.5 — a 397-billion-parameter sparse mixture-of-experts model with 17B active parameters, a 1M-token context window, and a small-model family the open-source community had been waiting for — they fired the person who built it. This is the story of what happens when a corporate open-source champion builds something so good it undermines the product his employer is trying to sell.

Every major language model in production today — GPT, Claude, Gemini, Llama — generates text the same way: left to right, one token at a time. That sequential assumption has been so productive for so long that most researchers treat it as fixed. A team at UC Berkeley and the University of Illinois just published dLLM: Simple Diffusion Language Modeling, a unified open-source framework that refuses to take autoregression for granted. Diffusion language models generate entire sequences through iterative denoising — bidirectionally, in parallel — and dLLM is the infrastructure that lets the field measure, compare, and build on them systematically for the first time.

As AI agents run for hundreds of turns with ninety-five percent KV-cache hit rates, the bottleneck shifts from compute to storage I/O. DualPath from Peking University, Tsinghua, and DeepSeek exploits idle decode-engine storage NICs to load KV-cache via RDMA, achieving nearly two times throughput on the same hardware. We break down the architecture, walk up the hardware ladder from Raspberry Pi clusters to DGX Spark rigs, and show that the minimum viable DualPath setup is eight Sparks with two switches for about thirty thousand dollars.

Someone finally ran a proper pentest on autonomous AI agents. Natalie Shapira, David Bau, and thirty researchers deployed LLM agents with persistent memory, email, Discord, and shell access then spent two weeks red-teaming them. Eleven failure modes, every one mapping to a known vulnerability class. We walk through the findings as engineering results, then pivot to EMPO-squared, a hybrid RL framework from Microsoft Research that trains agents to explore alternatives instead of defaulting to catastrophic compliance.

The most important robotics breakthrough of the last three years was not a new algorithm or a bigger model. It was making the hardware cheap enough to collect enough data. We trace the ALOHA lineage from a twenty thousand dollar bimanual teleoperation rig in a Stanford garage to Google DeepMind Gemini Robotics foundation model, following Tony Z. Zhao, Zipeng Fu, and Chelsea Finn across six papers and three years of compounding insight.

World ID's proof-of-personhood system went from a centralized iris database to a quantum-secure, open-source cryptographic protocol where no single entity holds biometric data. We walk through the Daugman iris code, Shamir Secret Sharing, Secure Multi-Party Computation, Anonymous MPC, zero-knowledge proofs, and why the math works even when the company behind it screwed up the rollout.

A team at Tsinghua University claims to have identified the specific neurons that predict when a large language model is about to hallucinate. Less than 0.1% of MLP neurons, identified via sparse logistic regression, generalize across domains and even detect fabricated entities with up to 97% accuracy. Most provocatively, amplifying these 'H-Neurons' increases sycophancy and compliance with harmful instructions — suggesting the capacity for over-compliance is baked in during pre-training, not alignment.

The FDA has cleared the first-ever human trial of a therapy designed to partially reverse cellular aging. Life Biosciences' ER-100, an epigenetic reprogramming treatment using a subset of Yamanaka factors delivered via AAV vector, will be injected into the eyes of patients with serious vision loss. We trace the science from Shinya Yamanaka's Nobel Prize to David Sinclair's controversial legacy, and ask whether partial reprogramming is the breakthrough or another cycle of premature optimism.

Microsoft Research published a complete system for writing data into borosilicate glass using femtosecond lasers. A palm-sized square holds nearly 5TB and survives for over 10,000 years. This episode traces the 30-year journey from Eric Mazur to Project Silica.

MIT researchers propose compressing LLM context in latent space rather than token space. Using closed-form linear algebra instead of gradient descent, Attention Matching achieves 50x KV cache compression in seconds — dramatically outperforming summarization on information-dense tasks like medical records QA. We cover the memory wall, why summarization fails, the three-step attention matching algorithm, nonuniform head budgets, and what this means for serving long-context models at scale.

The only full textbook on Ilya Sutskever's famous reading list. Why did a deep learning pioneer tell John Carmack to study algorithmic randomness? Because compression is intelligence — and this book is the mathematical foundation for that claim. We cover Kolmogorov complexity, the invariance theorem, incompressibility, algorithmic randomness, Berry's paradox, connections to Gödel and Turing, and why all of this matters for understanding why neural networks generalize.

NVIDIA introduces DreamZero, a 14-billion parameter World Action Model that jointly predicts future video and robot actions from a video diffusion backbone. Unlike Vision-Language-Action models that fail on physically novel tasks, DreamZero achieves over 2x improvement on generalization benchmarks and enables zero-shot transfer to unseen tasks like untying shoelaces — suggesting that the path to better robot policies runs through better video generation.

Our first DeepMind Dispatch covers three papers: Aletheia — a system that generates and verifies mathematical proofs autonomously; advances in Hutter optimization for large-scale model training; and Lyria 3, DeepMind's latest music generation model. We break down the technical substance and what it means for the frontier.

We present BitDance, a scalable autoregressive (AR) image generator that predicts binary visual tokens instead of codebook indices. With high-entropy binary latents, BitDance lets each token represent up to 2^256 states, yielding a compact yet highly expressive discrete representation. Sampling from such a huge token space is difficult with standard classification. To resolve this, BitDance uses a binary diffusion head: instead of predicting an index with softmax, it employs continuous-space dif

SkillRL from UNC Chapel Hill achieves 89.9% on ALFWorld with a 7B model — beating GPT-4o by 41.9 points. The secret: distilling raw experience into compact, reusable skills instead of storing verbose trajectory memories.

We explore a bold new framework that rethinks reinforcement learning from the ground up — replacing reward maximization with belief updating, and asking whether AI agents should learn the way scientists do.

Alibaba DAMO Academy built a complete embodied AI system in six months — eyes, hands, imagination, unified brain — and open-sourced everything. Seven model checkpoints, Apache 2.0, zero gating. This is the story of RynnBrain.

In July 2024, a French nonprofit's open-source voice AI went viral for demanding human sacrifice mid-conversation. Seven months later, the same team used the same architecture to build a real-time speech translator that runs on your phone. This is the story of Kyutai Labs — how Moshi became the Blood God, how Hibiki became a universal translator, and why transparency beats secrecy every time.

In a 48-hour span, three Chinese AI labs independently released frontier-class open-weight models. Step 3.5 Flash from StepFun delivers frontier intelligence with just 11 billion active parameters. MiniMax M2.5 offers comparable performance at one-twentieth the cost of Western alternatives. And GLM-5 from Zhipu AI trained a 744-billion parameter model entirely on Huawei Ascend chips — zero NVIDIA hardware. We break down the architectures, the benchmarks, the researchers behind them, and what it means when the frontier becomes a public good.

Researchers from UC Berkeley, NVIDIA, and UT Austin introduce DreamDojo, a framework that teaches robots physical skills by learning from large-scale human videos. Instead of expensive robot-specific data, DreamDojo distills 5 years of human video into a generalist world model that runs in real time. We break down how it works, why the team composition matters, and what it means for the future of robotics.

Researchers from MIT and Harvard propose Drifting Models, a new paradigm for generative modeling that achieves state-of-the-art image generation in a single forward pass. Instead of iterating at inference time like diffusion models, Drifting Models evolve the generated distribution during training using an elegant attraction-repulsion mechanism. The result: one-step image generation with FID 1.54 on ImageNet 256x256, beating even multi-step diffusion models. From the lab of Kaiming He, the creator of ResNet.

In our inaugural episode, we dive deep into Attention Is All You Need — the 15-page paper from June 2017 that introduced the Transformer architecture and reshaped all of artificial intelligence. We break down how it works, why the title is a Beatles joke, and where all eight authors ended up — from Google Gemini to RNA therapeutics to blockchain.