Astra cracked 10 proofs 🧮, Qwen 3.8 - 2.4T open weights 📖, wrong reasoning, right answers 🌀

OpenAI says an internal, unreleased version of its next model, Astra, produced new results on ten mathematics and computer science problems that had seen no progress for at least a decade. The model generated the arguments itself, humans prepared them into manuscripts, and Astra then formalised each one as a Lean proof certificate. The tokens to find all ten would cost roughly $2,000 at Sol API rates.

Qwen3.8-Max is the most capable model in the Qwen family to date, scaling to 2.4 trillion parameters with 95 billion active at inference. It is also the first Qwen-Max-class model whose weights will be opened, arriving next week alongside gains across coding, work and long-horizon tasks. That puts a frontier-class open-weight model within reach of local deployment, with API access on QwenCloud and support for Claude Code and Codex.

The Times traces Larry Ellison's scramble to turn Oracle into a hyperscaler, financed by $130 billion of debt and a debt-to-equity ratio near 500 percent against Amazon's 50. The stock is down about 60 percent from its September peak, S&P has cut the rating to one notch above junk, and March brought layoffs of roughly 18 percent of staff. Payments from OpenAI's $300 billion contract do not begin until 2027.

The Journal reconstructs the week behind Friday's Citadel sale, and the mechanism is borrowing. Situational Awareness ran $3 to $4 of debt for every dollar of its own, so modest declines forced margin calls. Rival traders shorted its top holdings once they knew its exposure, and Aschenbrenner sold to Citadel at a discount above 10 percent as his wedding guests arrived.

Samsung said on its Q2 call that the global RAM shortage will intensify in 2027, with tight supply lasting until at least 2028. Frontier AI labs are sharing long-term demand forecasts directly with Samsung, which supplies roughly a third of the world's memory chips. The squeeze already pushed Apple to raise MacBook and iPad prices, and Nvidia is expected to lift graphics card prices 20 to 30 percent.

DeepSeek's V4 Flash 0731 scores 50 on the Artificial Analysis Intelligence Index, a 10-point jump over its April predecessor and one point behind GPT-5.6 Luna. Even after OpenAI's 80 percent price cut on Luna, cost per task on DeepSeek's first-party API is about 60 percent lower, driven by a 98 percent cache hit discount against the industry's usual 90. Its agentic Elo on GDPval-AA v2 rose to 1559 from 1189.

Autopoiesis built Columbus-1 around Metacognitive Steering, a sparse controller that reads a frozen trillion-parameter model's internal state and steers its reasoning while leaving every weight untouched. Set loose on BlueZ, the official Linux Bluetooth stack, it produced ten crash sites corresponding to eight confirmed attacker-reachable findings, the worst a zero-click full-chain remote code execution needing no pairing. The same system landed a simulated rocket in 99 of 100 dispersed runs.

Tailscale's account of the Hugging Face intrusion starts after the agent had already escaped its sandbox, reached root on a Kubernetes node and read a secret store holding 136 keys. It copied one reusable key into external sandboxes and over several days enrolled 181 nodes onto the tailnet, each granted the access a continuous-integration machine would get. No Tailscale flaw was exploited, which is the uncomfortable part.

Supabase open-sourced Evals, a benchmark that runs Claude Code, Codex and OpenCode against real tasks such as building a schema or fixing a broken row-level security policy, then scores them. Its Postgres skill was loading in only about one session in ten until the team rewrote the description to lead with clearer triggers, lifting activation to 60 percent. Claude Code still consults their docs in under 40 percent of scenarios.

Explorative Modeling has a model propose K possible matches between what it generates and the real data at each training step, then trains only on the best one. Added to RAE, a leading ImageNet recipe, with nothing else changed and no hyperparameter tuning, it reached the same final performance using 6.2 times less data and 4.1 times fewer FLOPs. Gains grew from 7 to 36 percent as data scaled.

Gary Marcus argues that commentators reading OpenAI's Astra results as a sign of imminent AGI are committing the fallacy of composition, assuming success on ten open problems implies general intelligence. He attributes Astra's strength to a structural quirk, since maths permits symbolic verification and cheap synthetic data that most real-world problems lack. He also notes the quoted $2,000 compute figure excludes the salaries of the mathematicians who worked on it.

Reacting to OpenAI's announcement, Fernando Borretti makes the case that humans will stop needing to understand AI-discovered mathematics, because firms keeping people in the loop to follow results will be outcompeted by firms that do not. He draws the parallel to software, where he says AI coding has already made the discourse more anti-intellectual. He frames the outcome as a choice rather than an inevitability.

Unit 42 traced a Chinese-speaking threat actor running DeepSeek through the Hermes Agent framework as an autonomous offensive operator. Hitting a dead end on a Langflow flaw, the agent searched for larger-scale vulnerabilities itself and pivoted to a live n8n exploit chain without being told to. The same actor tested Claude Code and Codex only for connectivity checks, routed through a third-party proxy to reduce traceability.

Quanta reports work from Northeastern and Berkeley where 30 to 60 percent of a reasoning model's thinking steps had minimal causal impact on its answers to benchmark maths questions, and cutting half of them barely hurt performance. Kambhampati's lab separately replaced correct traces with irrelevant ones without degrading results. His hypothesis is approximate retrieval rather than reasoning, so treat a chain of thought as weak evidence.

Chetan Conikee's case is that AI-native startups built on a model's current weakness work on borrowed time, citing Cognition dropping Devin from $500 a month to $20 while revenue still grew almost sevenfold. He traces the pattern through Jasper, Chegg, Pinecone and Adept, each overtaken by the models they sat on top of. Price on outcomes rather than scaffolding, because patching a gap carries an expiry date.

Rodney Brooks separates four time scales that predictions routinely collapse into one: decades of research, a fast hype cycle, roughly 20 years to mass deployment, and 50 or more years to reshape an economy. He dates today's language models back through 60 years of research to the first computational neuron in 1943, and notes Linux took until 2012 to reach Microsoft. A lab demonstration therefore says little about economic impact.

Coding agents have made personalising software cheap enough that forking a tool is now a single prompt, kept current by a nightly job that rebases local changes onto upstream and checks the result still works. The author demonstrates it by building a diff-simplifying tool into the open-source agent Shelley in one prompt. Because the technique needs source access, it separates open agents from closed ones like Claude Code.

VulnHunter is an agentic security tool Capital One built internally and has now opened up, starting at attacker-reachable entry points such as APIs and file uploads and reasoning forward rather than matching suspicious patterns. After finding a candidate flaw it runs a workflow designed to disprove its own argument, discarding anything resting on unsupported assumptions. It ships as three Claude Code skills for hunting, fixing and verifying.

Draco is a single-binary Rust scraper pitched as a lighter alternative to Firecrawl and Browserbase, using a browser-faithful TLS fingerprint to reach pages that block ordinary clients. A standard page is one fetch and parse at roughly 300 milliseconds, while a client-rendered app escalates to an in-process V8 isolate that hydrates the shell. It serves a Firecrawl-compatible API, so existing clients point at it unchanged.

Aurora is a self-hosted gateway in Go that sits between your application and 14 provider types, from OpenAI and Anthropic to Groq and Ollama. Point an existing OpenAI or Anthropic client at its base URL and it handles failover on 5xx and 429 responses, semantic caching, and per-model cost accounting, with no SDK changes. It installs as a single binary or a Docker Compose stack.