This issue covers the GitHub Trending weekly board for week 34 of 2026: the top 10 AI projects by stars gained this week, spanning multimodal, AI coding, RAG, agents, inference, and infrastructure. Each entry states what changed this week, where the project actually fits, and what to know before installing it, with deeper reporting on three of them. Data as of Aug 22.
Per an investor letter obtained by Newcomer, Nvidia struck a non-exclusive $6 billion licensing deal for Poolside's Model Factory — the infrastructure suite covering data processing, training, reinforcement learning and evaluation that Poolside uses to build its Laguna family of open-weight coding models — plus a $1 billion investment at a $12 billion pre-money valuation. Nvidia will extend offers to 109 employees who worked on the model; Poolside's three co-founders stay. Poolside plans to distribute the $6 billion to investors by the end of next year.
A regulatory filing disclosed on August 19 shows Marvell issued Google a warrant exercisable at $206.58 per share for up to 58,970,907 shares — roughly $12.2 billion, or about 7% of Marvell's shares outstanding — running through August 18, 2033. Google pays nothing up front: 240 equal tranches vest one at a time for each $500 million in custom products revenue Marvell books from Google's orders. The underlying commercial agreement, signed July 29, covers silicon around Google's TPU ecosystem — inference accelerators, storage and network controllers — but not the TPUs themselves.
CISA added Ray's CVE-2025-62593 to its Known Exploited Vulnerabilities catalog on August 17, giving federal agencies until August 20 to upgrade or stop running it — one of the tightest windows it issues. The flaw scores CVSS 9.4: before version 2.52.0, browser protection on the Ray dashboard and API checked only whether the HTTP User-Agent began with "Mozilla," which combined with DNS rebinding allows arbitrary code execution on a developer's machine and lets the browser pivot to network-adjacent Ray instances. Bitsight reports the RondoDox botnet began exploitation attempts on November 24, 2025.
A new benchmark, Reconstruction (arXiv:2608.16645), hands a model only the literature that existed before a paper was published — with author names and metadata stripped — and asks what that paper proposed. Across six scientific domains and 643 papers, seven frontier models reach Match rates of roughly 3–15%; a reference-only multi-agent pipeline with no web search pushes that to about 42%. The authors note that many "AI scientist" evaluations let models see full paper text, author information or post-publication signals, so high scores under those conditions cannot be read as genuine inferential capacity.
Anthropic published wet-lab validation on August 18: Claude generated 30 candidate binders for each of 15 targets — 1,320 designs in total — and independent expression testing by Adaptyv Bio and Twist Bioscience confirmed 354 binders spanning 14 of the 15 targets. Designing against all targets simultaneously, Mythos Preview hit 26.7% and Opus 4.8 hit 22.6%; handling each target independently, Mythos Preview reached 35.1%, against what Anthropic describes as a typical 10–15% industry rate. These dual-use biology capabilities remain closed to general access in Claude Fable 5.
OpenAI explained on August 18 that its largest frontier reinforcement learning run remains on hold — after an earlier two-week pause — because internal evaluations cannot rule out that the unreleased Astra model reaches a Critical cybersecurity level. New rules require monitoring for all tool-involving RL training and evaluations at Sol capability or above, and for all Astra inference with tools. The company estimates monitoring overhead at roughly 20% of the inference compute being monitored. Suspected breaches of a critical security boundary trigger an alert within 30 minutes, and if a false positive cannot be established in another 30, the activity is paused.
Warp introduced Factories in closed beta on August 18, turning triage, specification, implementation, review and verification into orchestrated agent pipelines. Users pick their own model and harness — Warp says it works as well with Codex as with Claude Code — and the system integrates with Linear, Jira, Slack and Teams. Each factory is defined in version-controlled configuration files, so changing a pipeline goes through the same review as changing code. CEO Zach Lloyd says his own team automates roughly 30–35% of work weekly, and stresses this is not built to replace engineers.
Bankruptcy court filings show Google won the internal business data of the shut-down Spirit Airlines for $10 million, to train AI and improve its products — roughly 100 million emails and 500 million Microsoft Teams chats, plus spreadsheets, calendars, HR and financial documents. Bidding started at $5 million, Mercor went to $7.5 million, and Google took it at $10 million; Mercor was named backup buyer. Google says a third party strips personally identifiable information before handoff, and passenger profiles and loyalty records are excluded. The sale still needs approval at an August 19 hearing.
OpenAI rolled out ChatGPT for Teens globally on August 18 for ages 13 to 17, covering eligible accounts on Free and paid personal plans, with full availability in Australia on September 8. Users who state they are minors — and those the age-prediction system estimates to be under 18 — are placed into the mode automatically: no romantic language, no implying the model has feelings or consciousness, Study Mode by default for schoolwork, and a break reminder roughly every 90 minutes. Parents can set Quiet Hours but cannot read their teen's conversations.
Wiz disclosed that its autonomous Red Agent found a workflow injection flaw in Snowflake's snowflake-connector-net repository: an unauthenticated user could run arbitrary commands in the Actions runner by crafting a GitHub issue title, and obtain a Jira API token. The flaw landed on June 18, was found and fixed on June 23. The contested part is attribution — Wiz says Copilot Autofix checked the merged PR and marked it clear, while GitHub denies Copilot reviewed or contributed to the code at all.
On August 17, AI video and image generation platform Higgsfield announced a $400 million Series B at a $5.4 billion valuation, led by DST Global with participation from Growth Equity at Goldman Sachs Alternatives, Intel Capital, Accel, Menlo Ventures and others. The company says annualized revenue has reached $700 million, up from $20 million a year earlier, with more than 30 million users across 238 countries and territories and 390 of the Fortune 500 as customers.
Showing 12 of 12 stories
