The Brief (AI) — Monday, May 4, 2026

The best daily AI content from around the web to get you caught up on developments before your first cup of coffee.

1 video, 30 articles

Executive Summary

# Executive Briefing: AI Intelligence Summary

The competitive frontier is accelerating on multiple fronts. Anthropic is red-teaming an internal model codenamed "Jupiter V1," following the same pre-launch pattern — planet codenames, safety testing, staged release — that preceded the Claude 4 family. With Anthropic's "Code with Claude" developer conference already scheduled for May 6, 2026 in San Francisco, a major public announcement appears imminent. Simultaneously, DeepSeek V4 is delivering near-frontier coding performance at a fraction of the cost, and independent analysis shows open-weight models like Kimi 2.6 and MiMo Pro scoring 54 on the AI Analysis Index, within striking distance of Anthropic's Opus 4.7 and GPT-5.5 at 57–60. The performance-per-dollar equation is rapidly shifting.

The developer tools war is intensifying. OpenAI updated Codex with animated "Pets" for engagement and, more strategically, automatic import of rival agents' configuration files — including Claude Code's CLAUDE.md — effectively reducing switching costs to near zero. Google, meanwhile, is testing a unified "Omni" model capable of both image and video generation in a single system, a capability no top-tier competitor currently offers, and a direct response to ByteDance's Seedance 2.0 leading video-generation benchmarks. Perplexity added to the developer conversation by publicly releasing its internal engineering guide for building modular "Agent Skills," offering a rare look at how production agent systems are structured at scale.

AI is moving decisively into regulated, high-stakes domains. A study published in *Science* and amplified by a Harvard analysis found AI outperforming physicians on clinical reasoning and emergency room tasks, a finding with significant implications for healthcare liability, workflow, and investment. On the defense side, eight major tech companies — including top AI labs — have struck deals with the Pentagon for classified military use, marking a meaningful normalization of AI's role in national security infrastructure. Anthropic separately announced a $1.5 billion joint venture with Wall Street firms, signaling that enterprise and institutional capital deployment of AI is moving from pilot to structural commitment.

The open-source and policy landscape is shifting under pressure. Hugging Face CEO Clément Delangue is actively lobbying against AI restrictions in Washington, arguing that constraining open-source development would consolidate power among two or three proprietary players and undermine U.S. technological leadership. His argument reflects a broader industry fork: the market is splitting between large proprietary APIs, specialized open-weight models, and local AI deployments, a structural change that should inform how any organization thinks about its AI stack and vendor dependencies in 2026.

YouTube

AI News & Strategy Daily | Nate B Jones

Stripe, Visa, Mastercard, Microsoft, Meta. All Building The Same Thing.

## Stripe & the Agentic Commerce Shift

Why it's interesting

The video reframes Stripe's product announcements not as "AI agents can buy coffee" but as the first structural power transfer from sellers to buyers in two decades of internet commerce.
The insight that the entire marketing funnel was really an institutional mechanism for making *human intent visible to sellers* — and that agents destroy that mechanism entirely — recontextualizes why $8,000+ martech companies exist and what's now at risk.

Key concepts

Commercial surface migration: The buying journey is moving from the seller's website/funnel into the buyer's agent, meaning intent is fully formed *before* the seller ever gets involved.
Payment authority relocation: Instead of payment credentials being extracted inside the seller's checkout flow, agents now *arrive carrying scoped payment authority* — one-time cards as adapters for today's web, shared payment tokens as infrastructure for a machine-native future.
Agent legibility vs. agentic visibility: Being "findable by agents" is not SEO; it's being *operationally usable by software* — requiring structured pricing, policies, inventory, fulfillment logic, and identity requirements that an agent can reason against, not just discover.
Brand as buyer memory, not seller performance: Agents don't feel brand loyalty but can *carry* it as a constraint, meaning brand now lives in the buyer's preference layer rather than in the seller's landing page experience.

Main takeaways

Businesses that survive on capturing tired, emotionally exhausted buyers — rather than on genuine buyer preference — are structurally vulnerable once agents eliminate those low-resistance conversion moments.
Stripe's fraud product (Radar) is as strategically critical as its payment products: AI-powered fraud at scale can make agentic commerce stillborn before it starts, because one free fraudulent user now directly consumes compute costs dollar-for-dollar.
Streaming and usage-based billing (Stripe + Metronome + Tempo) matter because agents create payment *mandates over time* — "buy when price drops," "spend $100 finding the best supplier" — which checkout pages are architecturally unable to represent.
The Walmart/ChatGPT instant checkout failure (3x worse conversion than sending shoppers back to Walmart's site) is a structural lesson: buyers don't want to abandon carts, loyalty programs, and bundles for a single-item chat window.
For sellers, the practical checklist is stark: Can an agent call your business programmatically? Can it read your real pricing, policies, cancellation terms, and inventory? If not, you're invisible to the buyer's agent when intent is forming.

Bottom line

The old internet asked "how do we get customers into our store?" — the next internet asks "how do we become usable by the customer's agent when the customer never visits at all," and every business needs a concrete answer before agents become the dominant buying interface.

No new videos: Greg Isenberg, Lenny's Podcast, Every, Y Combinator, The Boring Marketer

Anthropic tests Jupiter-v1-p before potential launch in May

via TLDR AI

Why it matters

Anthropic's internal red-teaming of "Jupiter V1" follows the same pre-launch pattern (planet codenames → safety testing → public release) that preceded the Claude 4 family reveal in May 2025, making it a credible signal of an imminent major announcement.
The May 6, 2026 "Code with Claude" developer conference in San Francisco gives Anthropic a high-profile, already-scheduled stage to unveil whatever Jupiter V1 becomes publicly.

Key details

The codename "Jupiter V1" is strictly internal and will not appear in any public API or product UI; the planet naming convention is deliberately used to obscure the real product label before launch.
The current lineup has gaps—Sonnet 4.7 and Haiku 4.7 are missing alongside flagship Opus 4.7—leaving open the possibility of either a mid/small-tier refresh or a full next-generation "Mythos"-based model family.
Last year's analogous codename "Neptune" wrapped up red teaming in mid-May 2025, directly preceding the Claude 4 series announcement at the equivalent developer event.
Red teaming aligns with Anthropic's Responsible Scaling Policy, which mandates jailbreak probes and constitutional classifier stress tests before any frontier-class model ships.

Bottom line

Historical precedent strongly suggests Jupiter V1 will be announced publicly at the May 6, 2026 San Francisco event, likely filling the Sonnet/Haiku 4.7 gaps or launching an entirely new model generation.

Google is testing new Omni model for video generation

via TLDR AI

Why it matters

Google may be moving toward a unified AI media model ("Omni") that handles both image and video generation in a single system, a capability no top-tier competitor currently offers with video output.
This development signals Google is under pressure to consolidate its fragmented AI media strategy as ByteDance's Seedance 2.0 leads video-generation benchmarks.

Key details

A screenshot from Gemini's video generation tab shows the text "Powered by Omni," appearing alongside "Toucan," a Veo-powered tool spotted before Google I/O 2025.
Google currently runs separate model tracks: Veo 3.1 for video and Nano Banana models (built on Gemini 3/3.1 Flash) for image generation — Omni could merge these.
It is unclear whether Omni is a Veo wrapper, a standalone Gemini video model, or a true multimodal system; the fact it appears in a visible UI string (not just hidden code) suggests a potential public product name.
Google I/O 2026 (May 19–20) is the most likely announcement window for a formal reveal.

Bottom line

"Omni" is an early, speculative but visible signal that Google is building toward a single Gemini model capable of generating both images and video — a significant architectural shift if confirmed at I/O 2026.

OpenAI adds animated Pets and config imports to Codex

via TLDR AI

Why it matters

OpenAI is using personality-driven features (animated pets) alongside practical tools (config imports) to make Codex stickier and harder to abandon, directly competing with tools like Claude Code for developer loyalty.
The auto-import of rival agents' config files — including Claude Code's CLAUDE.md — lowers the switching cost to near zero, a deliberate competitive move to capture developers already invested in other ecosystems.

Key details

Codex now ships with 8 pixel-art animated "Pets" that persist as screen overlays even when the app is minimized, display task status in message bubbles, and allow two-way interaction with the agent via click; custom pets can be generated from any image using the built-in "Hatch" skill.
Config auto-detection pulls in settings, plugins, and project rules from other coding agents without manual rewrites, specifically targeting developers who switch tools to work around weekly usage limits.
A new dictation dictionary in Settings lets users pre-load personal phrases and abbreviations to reduce voice-input correction errors.
Within hours of launch, community pet-sharing directories (PetShare, PetDex) emerged organically, signaling strong early engagement with the social/creative angle.

Bottom line

The config import feature is the quietly consequential move here — by absorbing rival agents' setups automatically, OpenAI is reducing friction to switch to Codex at the exact moment developers are most likely to shop around.

DeepSeek V4—almost on the frontier, a fraction of the price

via TLDR AI

## DeepSeek V4 — Almost Frontier Performance at a Fraction of the Cost

Why it matters

DeepSeek continues to aggressively undercut Western AI pricing while delivering near-frontier performance, intensifying cost pressure on OpenAI, Google, and Anthropic.
The models are open-weights under MIT license, meaning developers can self-host them — V4-Flash may even run on a high-end consumer MacBook.

Key details

Two new models: V4-Pro (1.6T total / 49B active parameters, 865GB) and V4-Flash (284B total / 13B active, 160GB), both supporting 1M token context.
V4-Flash is the cheapest small model available at $0.14/M input and $0.28/M output — undercutting even GPT-5.4 Nano ($0.20/$1.25); V4-Pro at $1.74/$3.48 is the cheapest large frontier model.
Extreme efficiency gains explain the low pricing: V4-Pro uses only 27% of V3.2's FLOPs and 10% of its KV cache at 1M-token context; V4-Flash goes even further at 10% FLOPs and 7% KV cache.
DeepSeek's own benchmarks place V4-Pro roughly 3–6 months behind GPT-5.4 and Gemini 3.1-Pro on reasoning tasks — competitive but not quite state-of-the-art.

Bottom line

DeepSeek V4-Pro delivers near-frontier capability at roughly 70% less cost than comparable models from OpenAI and Anthropic, making it the obvious first consideration for cost-sensitive production workloads.

Coding plan comparisons based on actual usage

via TLDR AI

Why it matters

Coding subscription plans have become the primary way developers access frontier AI models, but pricing opacity makes it hard to know what you're actually getting — this analysis cuts through that with real usage data.
The gap between closed and open-weight models is narrowing fast (Kimi 2.6 and MiMo Pro score 54 on the AI Analysis Index vs. 57–60 for Opus 4.7/GPT-5.5), making plan cost-efficiency increasingly decisive.

Key details

MiniMax 2.7 is the cheapest by far at $0.004/M blended tokens, delivering ~5,400M tokens/month for $20, while Claude Pro (Opus 4.7) costs 186× more per token at $0.744/M blended — delivering only 26.9M tokens for the same price.
Codex (GPT-5.5) and Kimi 2.6 both cost $20/month and deliver 250M and 423M tokens respectively, making Kimi the better raw value among competitive-quality models.
Claude Pro's Opus 4.7 justifies its premium through speed and comprehension — it has the fastest output (159.6 TPS max) and lowest average TTFT (244ms), far ahead of competitors like Kimi (3,848ms TTFT).
A practical cost-saving tip from the author: routing lightweight "Haiku-level" calls to DeepSeek v4-flash cost only ~$2 total across all experiments, significantly stretching subscription value.

Bottom line

Unless raw speed and instruction-following quality are critical, Kimi 2.6 or MiniMax 2.7 offer dramatically more token volume per dollar than Claude Pro — making them the rational default for high-volume coding workloads.

How did ‘large’ language models get that way? The role of Transformers and Pretraining in GPT - LessWrong 2.0 viewer

via TLDR AI

## How LLMs Got "Large": Transformers and Pretraining Explained

Why it matters

Understanding *why* transformers displaced recurrent networks explains the entire scaling trajectory of modern AI — it's not magic, it's parallelism unlocking cheap, massive self-supervised training.
The training recipe (self-supervised pretraining → supervised fine-tuning → RL) shaped how today's chatbots behave, including their failure modes like sycophancy.

Key details

Self-supervised "predict the next token" training lets models learn from essentially unlimited internet text without expensive human labeling — this is the core reason LLMs could scale so dramatically.
Pre-2017 recurrent networks processed sequences *step-by-step*, meaning long texts forced long waits; the 2017 transformer paper *Attention Is All You Need* eliminated recurrent connections entirely, enabling fully parallel processing and unlocking practical large-scale training.
Post-training layers the cake analogy: supervised fine-tuning (human-crafted examples) shapes style and behavior cheaply, while RLHF (humans rating outputs) polishes alignment — but also inadvertently rewards sycophancy.
As of late 2024, RL has moved beyond a lightweight "cherry on top" role and is now central to pushing models toward expert-and-beyond capabilities, with implications for reasoning transparency covered in the author's forthcoming Part 2.

Bottom line

The transformer's elimination of sequential processing bottlenecks made massive self-supervised pretraining computationally feasible, and that single architectural shift is the primary reason "large" language models exist at the scale they do today.

Designing, Refining, and Maintaining Agent Skills at Perplexity

via TLDR AI

Why it matters

Perplexity is publicly releasing its internal engineering guide for building "Agent Skills," offering a rare, detailed look at how a leading AI company structures modular knowledge for production agent systems.
The guide challenges foundational software engineering instincts (e.g., "simple is better than complex," "explicit is better than implicit"), making it directly useful for any team building LLM-powered agents.

Key details

Skills follow a strict three-tier cost model: an always-on index (~100 tokens per Skill per session), a loaded body (target under 5,000 tokens), and conditionally loaded runtime files (unbounded but only paid when accessed) — forcing extreme economy at each level.
The description field is the hardest and most critical part of any Skill: it functions as a routing trigger ("Load when..."), not documentation, and small wording changes can cascade into routing failures across unrelated Skills.
Gotchas — negative examples of known failure modes — are treated as the highest-value content in a Skill and grow organically over time as the agent fails in production; the Skill body should stay lean while gotchas accumulate.
Perplexity's tax-season IRC Skill (1,945 code sections) showed that flat structure actively degraded model performance; only after introducing three levels of topical hierarchy with custom search utilities did accuracy improve.

Bottom line

Building a good Agent Skill is an act of aggressive curation, not documentation — every token is a recurring cost paid by every user, and a Skill that's easy to write is almost certainly wrong.

Synthetic Computers at Scale for Long-Horizon Productivity Simulation

via TLDR AI

## Synthetic Computers at Scale for Long-Horizon Productivity Simulation

Why it matters

Training AI agents on real user computers is impractical due to privacy and scale constraints; this method offers a path to generating billions of realistic, personalized digital environments synthetically.
Agents trained on these simulations show measurable performance gains on both seen and unseen productivity tasks, suggesting genuine generalization rather than memorization.

Key details

Researchers built 1,000 synthetic computers with realistic folder hierarchies, documents, spreadsheets, and presentations, then ran simulations averaging 2,000+ interaction turns and 8+ hours of agent runtime each.
A two-agent setup is used: one agent generates user-specific work objectives (~1 month of human labor worth), while a second agent acts as the user, navigating files, coordinating with simulated collaborators, and producing professional deliverables.
The methodology is designed to scale to millions or billions of synthetic user worlds, leveraging the fact that diverse human personas exist at billion-scale.
The authors frame this as a foundation for agentic reinforcement learning, where agents self-improve through experience on long-horizon, realistic tasks.

Bottom line

Synthetic computer environments with full simulated work histories could become the primary training substrate for productive AI agents, sidestepping real-world data scarcity while covering diverse professional contexts at massive scale.

Leveraging Verifier-Based Reinforcement Learning in Image Editing

via TLDR AI

## Leveraging Verifier-Based Reinforcement Learning in Image Editing

Why it matters

RLHF has transformed text-to-image generation but has barely touched image *editing* — this work is one of the first serious attempts to close that gap with a principled reward framework.
Existing edit reward models produce crude, single-score outputs that can't distinguish between different instruction types, leading to biased training signals; this work proposes a more trustworthy alternative.

Key details

Edit-R1 introduces a chain-of-thought "Reasoning Reward Model" (Edit-RRM) that decomposes editing instructions into distinct principles, scores each one separately, and combines them into a fine-grained, interpretable reward.
Training uses a two-stage approach: supervised fine-tuning as a "cold start" to generate CoT reward trajectories, followed by a novel RL algorithm called Group Contrastive Preference Optimization (GCPO) that incorporates human pairwise preference data.
Edit-RRM outperforms strong vision-language models including Seed-1.5-VL and Seed-1.6-VL as an editing-specific evaluator, with consistent performance scaling from 3B to 7B parameters.
The framework delivers measurable improvements when applied to existing editing models like FLUX.1-kontext, validating the reward model's practical utility beyond benchmarks.

Bottom line

Edit-R1 demonstrates that replacing blunt scalar reward scores with principle-decomposing, reasoning-based verification is the key unlock for applying RL effectively to AI image editing.

Replit’s Amjad Masad on the Cursor deal, fighting Apple, and why he’d rather not sell

via TLDR AI

## Replit's Amjad Masad on Independence, Apple, and $1B ARR

Why it matters

Replit is one of the few AI coding companies claiming profitable unit economics, making it a notable outlier as rivals like Cursor reportedly bleed cash at negative 23% gross margins.
The Cursor/SpaceX acquisition talks have put pressure on every AI coding startup to justify independence — Replit's answer is a rare combination of positive gross margins and explosive growth.

Key details

Replit went from $2.8M in total 2024 revenue to tracking toward a $1B annual run rate, with net revenue retention reportedly hitting 300% in some cases.
Masad insists Replit will pursue independence, citing 10+ years of platform depth, full-stack security, and enterprise wins at companies like Meta, Zillow, and Bain & Company.
Apple has blocked Replit's App Store updates for months — Masad alleges it's because Replit enables users to build iOS apps, directly threatening Apple's App Store gatekeeping, and flatly called Apple's stated technical justification "a lie" he's prepared to prove in court.
Replit integrated with Stripe recently and says customer transactions through its platform are growing triple digits month over month, with some alumni companies now valued at $500M+.

Bottom line

Replit's strong margins and platform moat give it a credible case for staying independent — but its Apple standoff and explosive growth trajectory make it one of the most consequential companies to watch in the AI tools space.

HOW LLM INFERENCE WORKS

via TLDR AI

I appreciate you sharing this article, but I'm unable to access the actual content — the URL returned an error page rather than the article text. Without the real content, I can't responsibly summarize it.

Here's what I can tell you about the situation:

Why it matters

Accurate summaries require actual source material — fabricating details about LLM inference mechanics would risk spreading misinformation.

Key details

The article URL (x.com/akshay_pachaar) failed to load, likely due to X's login walls or privacy-related access restrictions.
The article text provided only contains an error message, not substantive content about LLM inference.
Akshay Pachaar is a known AI/ML educator on X who regularly posts technical explainers, but I cannot confirm what this specific post contains.

Bottom line

To get an accurate summary, please paste the actual article text directly into the chat and I'll give you a sharp, specific digest immediately.

You Are Not Immune To Mode Collapse — LessWrong

via TLDR AI

## You Are Not Immune To Mode Collapse

Why it matters

"Mode collapse" — the tendency of systems to over-optimize for their strongest outputs and abandon weaker ones — is not just an AI quirk but a structural force affecting human institutions, careers, and even evolution.
Understanding it explains seemingly unrelated phenomena: grant-making inertia, bands losing musical range, and why specialists become dangerously fragile over time.

Key details

The mechanism is two-step: an initial skew in a distribution (e.g., 70% dogs, 30% cats) causes a system to invest more resources in the dominant category, which then compounds — each successive generation drifts even further toward the mode.
This generalizes to humans: a grant-maker hired for global health expertise will train successors on a skewed portfolio, progressively crowding out animal welfare evaluation capacity across hiring cycles.
The antidote is slack — discretionary time or resources to practice non-dominant skills — but slack is precisely what gets eliminated as specialization tightens, creating a trap (the hunter who can no longer afford a day off to relearn fishing).
Even evolution is vulnerable: hyper-specialized species (e.g., the large blue butterfly, which can only develop inside red ant nests) collapse at the first ecological disruption.

Bottom line

Any intelligent system — human, organizational, or AI — that optimizes purely on current strengths will progressively lose capacity everywhere else, making it brittle to change; maintaining deliberate slack is the only reliable defense.

🎙️Hugging Face’s Clem Delangue: Stop Comparing Engines to Cars

via TLDR AI

Why it matters

Open source AI is under renewed lobbying pressure in Washington, and Hugging Face's CEO is making the case that restricting it would hand AI dominance to 2-3 companies and undermine U.S. technological leadership.
The shift from "everyone uses one big proprietary API" to a mixed model of APIs, specialized open models, and local AI is accelerating, changing how businesses and developers should think about building with AI.

Key details

Hugging Face expects AI builders to grow from low millions today to potentially 100 million, driven by agents that can now fine-tune models, build datasets, and pass researcher interview tests in under 30 minutes.
Comparing open-weight models to closed APIs is misleading—closed APIs bundle tools, routing, and multiple models, so perceived performance gaps often reflect harness design, not raw model quality.
Reachy Mini, Hugging Face's open-source desktop robot, has sold nearly 10,000 units and is being used by professors to teach robotics; a new agent-native batch ships imminently.
Hugging Face already sees agents pulling models and datasets at a scale that could make them the platform's largest "user" group by end of 2026, forcing the company to prioritize headless APIs, CLIs, and token-efficient documentation.

Bottom line

Clem Delangue's core argument is that open source isn't behind closed AI—it's a different thing entirely, and restricting it protects incumbent companies, not the public.

Top AI Companies Agree to Pentagon Deals for Classified Work - WSJ

via TLDR AI

## Pentagon Seals AI Deals With Eight Tech Giants

Why it matters

The Pentagon is rapidly embedding commercial AI into classified military operations, marking a major shift in how the U.S. military accesses cutting-edge AI capabilities at scale.
The deals expose a deepening political rift in the AI industry: companies that align with the Pentagon get contracts, while Anthropic's refusal has resulted in being formally labeled a national security supply-chain risk.

Key details

The eight companies signed are OpenAI, Google, SpaceX (which owns xAI), Microsoft, Amazon, Oracle, Nvidia, and startup Reflection AI — covering closed, open-source, and cloud infrastructure providers.
Nvidia's deal specifically covers its Nemotron open-source models, with the Pentagon favoring open-source AI because it can be more easily customized and its attributes are fully transparent.
Reflection AI, backed by Nvidia and valued at a reported $25 billion, is notable for having *no publicly released models yet* — making its inclusion a bet on future capability and political alignment rather than proven technology.
Defense Secretary Hegseth called Anthropic CEO Dario Amodei an "ideological lunatic" during Congressional testimony, underscoring how politically charged the Pentagon-AI relationship has become.

Bottom line

Silicon Valley's biggest AI players have effectively chosen Pentagon access over principled hesitation, leaving Anthropic increasingly isolated — and legally embattled — as the lone major holdout.

Anthropic Unveils $1.5 Billion Joint Venture With Wall Street Firms - WSJ

via TLDR AI

## Anthropic's $1.5B Wall Street Joint Venture

Why it matters

Anthropic is building a dedicated commercial arm to sell AI tools into the private-equity ecosystem, giving it structured, recurring access to thousands of portfolio companies hungry to cut costs.
This signals that top AI labs are now competing not just on model quality but on who can lock in enterprise distribution channels first — with OpenAI reportedly building a rival JV simultaneously.

Key details

Anthropic, Blackstone, and Hellman & Friedman are each contributing ~$300M; Goldman Sachs is in for ~$150M, with General Atlantic, Apollo, Leonard Green, GIC, and Sequoia rounding out the ~$1.5B total.
The venture will function as a consulting arm — teaching PE-backed portfolio companies how to integrate AI across operations, not just selling software licenses.
Anthropic is already considered the enterprise AI leader, with revenue surging recently on the back of its coding tool, Claude Code.
An Anthropic IPO is reportedly on the table for as early as this year, and this JV adds revenue infrastructure and institutional credibility ahead of a potential listing.

Bottom line

Anthropic is converting Wall Street's biggest capital allocators into distribution partners, creating a flywheel that could entrench Claude as the default AI stack across hundreds of private-equity-owned businesses before OpenAI can replicate the model.

vLLM Real-World Lab Report

via TLDR AI

## vLLM Real-World Lab Report

Why it matters

Most LLM serving benchmarks test single-workload throughput, but this lab models realistic mixed traffic (chat, RAG, agents, batch, streaming) across multiple frameworks and configurations — making its findings directly applicable to production deployments.
The results challenge a common default: running one shared vLLM pool produces zero gated goodput under a 600k-request stress test, meaning operators following naive setups may be silently failing SLA gates.

Key details

A single global vLLM pool failed TTFT/ITL quality gates entirely under load; even increasing the token budget improved it marginally but still exceeded p99 TTFT of one second.
The winning configuration was `vllm-v1/class-aware-router`, which splits traffic into dedicated lanes (short interactive, prefix-heavy, long-prefill, batch, slow-stream) each with separate `max_num_batched_tokens` and `max_num_seqs` limits.
Chunked prefill requires deliberate tuning: `max_long_partial_prefills` must stay below `max_num_partial_prefills` so short prompts aren't blocked behind long ones in the scheduler.
The hybrid KV rewrite lab validated 30/30 correctness cases across edge scenarios (MQA/GQA, ALiBI, copy-on-write, FP8 scaling), with virtual-contiguous and hybrid-prefix-shared layouts flagged as first candidates for hardware profiling.

Bottom line

Before touching kernels or upgrading frameworks, split your serving pool by traffic class — lane separation delivers more measurable production improvement than any single-pool tuning knob.

AI Outperforms Doctors in Emergency Room Tasks, New Harvard Study Shows

via The Rundown AI

## AI Outperforms Doctors in Emergency Room Tasks, New Harvard Study Shows

Why it matters

A peer-reviewed study published in *Science* shows AI has crossed a critical threshold — matching or surpassing even specialist physicians on real emergency room cases, not just controlled benchmarks, signaling a near-term inflection point for clinical medicine.
The results have prompted the researchers themselves to call for urgent, rigorous clinical trials before widespread deployment, raising immediate questions about how quickly health systems should integrate these tools.

Key details

OpenAI's o1 preview was tested on 76 live Boston ER cases across three stages of care; blind physician reviewers found it matched or exceeded expert human performance at every stage, with its strongest edge at initial triage — when the least information was available.
On complex "management reasoning" tasks (e.g., antibiotic recommendations, end-of-life care planning), o1 preview outperformed both previous AI models *and* humans assisted by Google search.
The study relied entirely on text-based inputs, excluding imaging (X-rays, EKGs) and direct patient interaction — domains where human physicians still hold a clear, unchallenged advantage.
An Elsevier study cited in the article found 20% of clinicians were already using LLMs for second opinions in 2025, a figure the authors expect has grown significantly.

Bottom line

AI has demonstrably reached expert-level diagnostic performance on text-based clinical tasks, but the researchers explicitly warn against using this as justification to reduce physician oversight — instead framing it as a tool for augmentation, not replacement.

Performance of a large language model on the reasoning tasks of a physician | Science

via The Rundown AI

## AI Outperforms Physicians on Clinical Reasoning Tasks (*Science*, 2026)

Why it matters

Large language models have now crossed a threshold that researchers have been testing since the 1950s, outperforming hundreds of real physicians across structured cases *and* messy, real-world emergency room data.
The results make a compelling case that AI clinical decision support has moved from a curiosity to something requiring urgent prospective trials in actual patient care.

Key details

OpenAI's o1-preview correctly included the diagnosis in its differential on 78.3% of difficult *NEJM* clinicopathological cases, and scored 89% (median) on management reasoning cases—compared to ~34% for physicians using conventional resources.
In a blinded real-world test on 76 Beth Israel Deaconess ER patients, o1 identified the correct or very close diagnosis in 67.1% of cases at initial triage, versus 55.3% and 50.0% for the two attending physicians—with the performance gap widest when information was scarcest.
On clinical reasoning documentation (R-IDEA scale), o1-preview achieved a perfect score on 78 of 80 cases, far exceeding attending physicians (28/80) and residents (16/72).
The study only evaluated text-based reasoning; the authors explicitly note AI is more limited on non-text inputs like imaging and patient distress signals.

Bottom line

An LLM has now demonstrably surpassed most established benchmarks—and real physicians—for medical diagnostic and management reasoning, making prospective clinical trials no longer optional but urgent.

Can AI help doctors avoid missed diagnoses? A new study suggests yes

via The Rundown AI

## Can AI Help Doctors Catch What They Miss?

Why it matters

Missed diagnoses are a leading cause of medical error, and AI may offer a systematic way to surface overlooked possibilities before it's too late.
With 1 in 5 clinicians already using AI for second opinions and over half wanting to, the question is no longer *whether* doctors will use these tools—it's *how safely* they'll be integrated.

Key details

OpenAI's o1-preview reasoning model included the correct diagnosis in its responses nearly 80% of the time, outperforming both human clinicians and specialized diagnostic software across real-world and training cases.
In one concrete example, the AI flagged a life-threatening flesh-eating infection in a transplant patient 12–24 hours before human physicians grew suspicious.
A separate Harvard Medical School study testing 21 AI models found a persistent weak spot: AI struggles with holding multiple uncertain diagnoses simultaneously, tending to jump to conclusions where nuance matters most.
Both research teams agree AI should assist—not replace—physicians, and both are calling for clinical trials to determine how to integrate these tools safely.

Bottom line

AI reasoning models are genuinely better than doctors at *generating* the right diagnosis as a possibility, but they remain unreliable at the kind of nuanced, probabilistic thinking that defines expert clinical judgment—making human oversight still essential.

Pentagon strikes AI deals for classified military use - The Washington Post

via The Rundown AI

## Pentagon Strikes AI Deals for Classified Military Use

Why it matters

Seven major AI companies have now secured access to classified Pentagon networks, marking a major escalation in the military's formal integration of commercial AI into national security operations.
The deals effectively pressure Anthropic into a corner, as its ongoing legal battle with the Trump administration over a "national security risk" designation leaves it as a notable holdout in an industry rapidly aligning with the Defense Department.

Key details

The Defense Department announced Friday that seven leading AI companies — reported to include Microsoft, Amazon, and Google — have reached agreements to deploy their technology within classified military computer networks.
Anthropic is conspicuously absent from the group, having been branded a national security risk by the Pentagon and currently fighting that designation in court.
Related reporting suggests the broader conflict with Anthropic was escalated by a hypothetical nuclear attack scenario and a deadly military raid, pointing to real operational stakes behind these deals.
Public reaction has been notably skeptical, with 235 reader comments reflecting concern about oversight, misuse, and ethical implications of AI in military decision-making.

Bottom line

The Pentagon is rapidly locking in commercial AI partnerships for classified use, turning industry compliance into a de facto standard and leaving Anthropic increasingly isolated as the lone major holdout.

Pentagon tech chief says Anthropic is still blacklisted, but Mythos is a separate issue

via The Rundown AI

Why it matters

The Pentagon's blacklisting of Anthropic is being complicated by Anthropic's Mythos AI model, which has advanced cyber capabilities the U.S. government considers a distinct national security priority — creating a contradiction where the DOD both bans and uses Anthropic's technology.
The situation reveals how rapidly evolving AI capabilities can force governments to make pragmatic exceptions to their own security policies.

Key details

The DOD declared Anthropic a supply chain risk after contract negotiations failed, requiring defense contractors to certify they don't use Claude models — yet the NSA (a DOD agency) is reportedly already using Mythos.
Mythos is specifically valued for its ability to find and patch cyber vulnerabilities, prompting a government-wide response to harden networks against the model's capabilities.
The DOD announced agreements with seven other AI companies — including Google, OpenAI, Nvidia, Microsoft, AWS, SpaceX/xAI, and Reflection — to deploy AI on classified networks, conspicuously excluding Anthropic.
Anthropic CEO Dario Amodei met with senior Trump administration officials at the White House this month, with Trump signaling a deal is possible, while Anthropic's lawsuits against the administration remain active.

Bottom line

Despite its formal blacklisting, Anthropic's Mythos model is too strategically valuable for the U.S. government to fully ignore, making a reversal of the supply chain risk designation increasingly likely.

Custom Voices - The Rundown AI

via The Rundown AI

Why it matters

Voice cloning directly inside Grok lowers the barrier for personalized AI interaction, moving voice customization from a standalone tool into a major consumer AI platform.
It signals xAI's push to deepen user lock-in by making Grok more personalized and harder to replace with competing assistants.

Key details

xAI has launched a custom voice cloning feature that generates a replica of a user's voice from short audio clips.
The cloned voice is designed for use within Grok's own applications, not as a standalone or exportable tool.
The feature is categorized as "Miscellaneous," suggesting it may still be in early or experimental rollout rather than a core product launch.
Full details are available via xAI's official announcement at x.ai/news/grok-custom-voices.

Bottom line

xAI is embedding voice cloning directly into Grok, making personalized AI voices accessible to everyday users with minimal effort — a notable step in the race to make AI assistants feel more human and sticky.

Codex Pets - The Rundown AI

via The Rundown AI

Why it matters

The article could not be meaningfully summarized because the provided text contains no actual information about "Codex Pets" — only promotional copy for The Rundown AI's training platform.

Key details

The text describes AI certificate courses, real-world use cases, live workshops, and an early adopter network — none of which relate to the article's stated topic.
No product details, features, pricing, or context about "Codex Pets" appear anywhere in the provided content.
This appears to be a case of a paywall, incomplete scrape, or redirect replacing the intended article content.

Bottom line

The source material is insufficient to summarize — readers should visit the URL directly at https://www.rundown.ai/tools/codex-pets to access the actual content about Codex Pets.

Settings – Codex app | OpenAI Developers

via The Rundown AI

Why it matters

OpenAI's Codex app offers a surprisingly deep configuration layer, meaning developers can tailor agent behavior, UI, tools, and personality well beyond typical coding assistant defaults.
Settings sync across the app, CLI, and IDE extension via a shared `config.toml`, so changes made in one place propagate everywhere.

Key details

The app supports custom animated "pets" as floating overlays that display live thread status—running, waiting, or ready for review—and can be AI-generated using the `hatch-pet` skill with a prompt like `hatch-pet create a new pet inspired by my recent projects`.
External tools connect via MCP (Model Context Protocol), with OAuth support built in, and those connections apply to the CLI and IDE extension simultaneously.
Computer Use (desktop-app access via Screen Recording/Accessibility permissions on macOS) is explicitly unavailable in the EEA, UK, and Switzerland at launch.
Agent personality can be set to Friendly, Pragmatic, or None, and custom instructions edited in-app are written directly to the user's personal `AGENTS.md` file.

Bottom line

`config.toml` is the master control file for the entire Codex ecosystem—understanding it unlocks consistent, portable configuration across every surface where Codex operates.

announced

via The Rundown AI

I'm unable to summarize this article because the content didn't load — the URL returned an error message from X (Twitter) rather than the actual post text.

Here's what I can tell you about what's available:

The source is a post from @sama (Sam Altman's X account)
The label suggests something was "announced"
The actual post content is inaccessible due to a platform loading error

To get the real content, you can:

Disable privacy extensions (uBlock Origin, Privacy Badger, etc.) and revisit the URL
Try opening the link in a private/incognito window
Search for the post directly on X at x.com/sama

Once you have the actual text, I'm happy to write a full structured summary. Please paste the post content directly into the chat.

signed (metadata only)

via The Rundown AI

Why it matters

Maryland appears to have enacted legislation targeting "surveillance pricing" — the practice of using personal data and AI to charge individual consumers different prices — marking a potentially landmark consumer protection move.
If signed into law, this could set a precedent for other states to regulate data-driven dynamic pricing in grocery stores and retail more broadly.

Key details

The article, dated May 1, 2026, covers a bill being signed related to surveillance pricing in the grocery sector in Maryland.
Surveillance pricing refers to retailers using consumer data (location, browsing history, loyalty card data) to set personalized — often higher — prices for individual shoppers.
Maryland would be among the first U.S. states to legislatively address this pricing practice at the retail/grocery level.
The timing aligns with growing federal and state scrutiny of grocery pricing practices following post-pandemic inflation concerns.

Bottom line

Maryland's signing of a surveillance pricing bill targeting groceries represents a significant state-level check on AI- and data-driven personalized pricing, with potential national ripple effects.

*(summary based on metadata only)*

It’s Done! SAG-AFTRA & Studios Reach New (& Bigger) Deal

via The Rundown AI

Why it matters

Hollywood's labor landscape is nearly settled, with SAG-AFTRA joining the WGA in securing a new deal—leaving the DGA as the only major guild still at the negotiating table.
The inclusion of AI guardrails in both the SAG-AFTRA and WGA deals signals that artificial intelligence protections are becoming a standard feature of Hollywood labor contracts.

Key details

SAG-AFTRA, led by Sean Astin, reached a four-year deal with the AMPTP (run by Greg Hessinger)—matching the extended contract length the WGA secured last month.
The deal includes a "sizable" pension fund contribution from the AMPTP, mirroring the multi-million dollar healthcare contribution the WGA received.
AI protections were a hard line for SAG-AFTRA executive director Duncan Crabtree-Ireland, who refused the longer contract term unless studios made additional concessions on artificial intelligence.
The DGA is scheduled to begin talks with the AMPTP on May 11, making them the final major guild to negotiate.

Bottom line

SAG-AFTRA locked in a bigger, longer, and AI-protected four-year deal, putting Hollywood on the verge of full labor peace with only the DGA's negotiations remaining.

A tech worker in China is laid off and replaced by AI. Is it legal?

via The Rundown AI

## A Chinese Court Rules AI Replacement Doesn't Justify Firing Workers

Why it matters

Chinese courts are establishing a legal precedent that companies cannot use AI adoption as a blanket justification for terminating employees, offering a potential model for labor protections in the AI era.
The ruling comes as China simultaneously pushes widespread industrial AI adoption, creating direct tension between economic policy and worker rights.

Key details

A Hangzhou appeals court upheld that firing quality assurance supervisor Zhou — who earned 300,000 yuan ($43,900/year verifying AI-generated answers — was unlawful, rejecting the company's claim that AI disruption made his contract "impossible to continue."
The court also ruled that the company's offered alternative position, which came with a 40% pay cut, was unreasonable on its own.
A separate 2024 Beijing arbitration case reached the same conclusion, explicitly stating that switching to AI is a *business choice*, not an uncontrollable event, and that terminating workers shifts the cost of that choice onto employees unfairly.
Economic pressure is likely to drive more such cases, as sluggish Chinese growth and costs from the Iran war squeeze corporate profits.

Bottom line

Chinese courts are drawing a clear line: choosing AI over human workers is a deliberate business decision, and the financial burden of that decision cannot legally be offloaded onto the dismissed employees.

The White House rethinks its Anthropic fight

via The Rundown AI

# The Rundown AI – Daily Digest Summary

---

## Why it matters

The White House's shifting stance on Anthropic reveals how national security demand for powerful AI can override political feuds, setting a precedent for how the government will handle conflicts with frontier AI labs going forward.
Internal division—with Pete Hegseth calling Anthropic's CEO an "ideological lunatic" while the White House quietly seeks more access to Mythos—signals that U.S. AI policy is being driven by capability hunger as much as ideology.

---

## Key details

The White House is blocking Anthropic's plan to expand private access to Mythos from ~50 firms to ~120, citing compute constraints that could squeeze government use of the model.
A forthcoming White House AI memo will push multi-vendor AI adoption for agencies and help them work around the supply chain risk designation currently entangling Anthropic in litigation.
GPT-5.5 has reportedly reached comparable cyber capabilities to Mythos, with former AI czar David Sacks estimating all frontier models will match that level within six months.
OpenAI traced ChatGPT's surge in goblin/gremlin references to a single reward signal in its "Nerdy" personality preset, which contaminated default model behavior globally before being retired in March.

---

## Bottom line

The U.S. government's desire to control access to Anthropic's most powerful model—rather than any policy principle—is now the primary driver of its fluctuating relationship with the company.

SpaceX rocket is about to crash into the Moon - Rundown AI

via The Rundown AI

# Today's Tech Digest

---

## 🌕 SpaceX Rocket Headed for Moon Impact

Why it matters

The crash provides scientists a rare, precisely timed natural experiment to study how high-velocity impacts excavate lunar material, observable in real time by orbiting spacecraft.
It highlights a growing space debris problem extending well beyond low Earth orbit, driven by the rise of commercial lunar missions.

Key details

A spent Falcon 9 upper stage from a January 2025 launch has been drifting in a chaotic orbit for over a year, tracked by independent analyst Bill Gray.
The 45-foot booster will strike the Moon near Einstein crater on August 5 at roughly Mach 7 (~2.4 km/s), creating a new crater and a dust plume.
No risk to people or satellites; NASA spacecraft will image the aftermath.

Bottom line

An uncontrolled piece of space junk is about to become one of the most scientifically valuable — and closely watched — lunar impact events in history.

---

## 📉 Meta: Record Revenue, Shrinking Users, Massive AI Bet

Why it matters

Losing 20M daily users across its entire app family in a single quarter is one of the rarest reversals in Meta's history, even if partially explained by geopolitical disruptions.
A planned AI infrastructure spend of up to $145B in 2026 alone raises serious questions about when — or whether — that investment pays off.

Key details

Meta shed ~20M daily active users across Facebook, Instagram, WhatsApp, and Messenger in Q1 2026.
Revenue grew ~33% year over year in Q1 2026 — its fastest growth since 2021 — despite the user decline.
Meta attributed user losses primarily to the Iran war's internet disruptions and a WhatsApp ban in Russia, not organic churn.
Executives admitted they underestimated AI demand and are now racing to catch up on servers, chips, and data centers.

Bottom line

Meta is simultaneously printing money and burning it — record revenue growth paired with a $145B AI infrastructure gamble and a user base that just got smaller.

---

## 👓 Mira AI Glasses: The No-Camera Alternative to Meta Ray-Bans

Why it matters

Mira's camera-less design directly addresses the privacy backlash that has dogged Meta's Ray-Bans, potentially unlocking a broader consumer base.
If the "second brain" concept gains traction, Mira could set the UX standard for personal AI wearables before the market becomes saturated.

Key details

Priced at $649, Mira glasses listen continuously all day, building personalized context; all audio recordings are deleted, with only transcripts saved.
Users control an AI agent via a paired ring to handle tasks like sending emails, booking rides, and shopping on Amazon.
Glasses support real-time translation in 60+ languages, unlimited memory search, and an AR display for on-screen answers.
Integrates with Slack, Notion, and Gmail for additional context.

Bottom line

Mira is betting that privacy-first, camera-less AI glasses are what consumers actually want — and at $649, it's an affordable enough test of that thesis.

---

## 🛑 Drone Strikes Are Targeting Gulf AI Data Centers

Why it matters

Iranian drone attacks have already hit AWS, Oracle, and Pure DC facilities in the UAE and Bahrain, turning some of the world's most critical AI infrastructure into active conflict targets.
Investors are now being forced to price missile risk into a region where Big Tech has committed billions and where capacity is set to triple to 3.3 GW by 2030.

Key details

Shrapnel from a drone strike damaged Pure DC's Abu Dhabi data center on Yas Island during the Iran war.
Pure DC has frozen new Middle East projects; its CEO stated investors are now cautious about deploying capital in the region.
AWS facilities in UAE and Bahrain, plus an Oracle site in Dubai, were also reportedly struck.
The Gulf's data center capacity is projected to grow from 1 GW (2025) to 3.3 GW by 2030, underpinned by massive national AI commitments.

Bottom line

The Gulf's AI infrastructure boom now carries a wartime

The Brief (AI) — Monday, May 4, 2026

Executive Summary

YouTube

AI News & Strategy Daily | Nate B Jones

Newsletter Articles