AI Daily · 2026-08-21

Today's standout is Poolside's reverse execuhire with Nvidia: Nvidia invests $1B at a $12B pre-money valuation, takes in roughly 109 employees, and signs a $6B non-exclusive licensing deal, while Poolside's founders stay and its infrastructure spinout keeps building a 1.2GW Texas data center on the way to 7GW neocloud. Other updates center on model efficiency and developer tooling: LiquidAI's DSpark draft models speed up LFM2.5 inference up to 3.2x via speculative decoding; LangSmith adds preview builds for testing agent changes before merge; ChatGPT search queries using site: jumped to 16-17% around the GPT-5.6 release, suggesting a retrieval change. Bun 1.4's WebView enables a zero-dependency screenshot API, and the /wayfinder skill helps navigate planning when the end state is unclear.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐ [Model Release] Up to 3.2x Faster Inference with LFM2.5-DSpark

LiquidAI released DSpark draft model checkpoints for three LFM2.5 models (LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B), enabling speculative decoding with up to 3.18x throughput on GPU and up to 2.87x on-device, and cutting function-calling latency by 57% on average for LFM2.5-2.6B. The draft models are roughly 300M parameters with a 5-layer attention-only design, trained on a larger, more diverse mix of SFT, chat, code, and function-calling data; the checkpoint with the highest acceptance rate was selected. Day-one support for llama.cpp and SGLang is included, with the DSpark integration open-sourced upstream. Because accepted tokens match the target model under greedy decoding, benchmark accuracy remains unchanged.

Product Update

⭐⭐ [Product Update] Test Agent Changes with LangSmith Preview Builds

LangSmith introduces Preview Builds, a new deployment feature for testing agent changes before merging. It creates a temporary, isolated staging environment from a pull request branch, letting reviewers test the proposed agent without cloning the repo. LangSmith automatically rebuilds the preview when new commits are pushed, keeping it in sync with the PR. Teams can trigger previews on every PR or only via a configured GitHub label, and can manage them with idle TTL, concurrent limits, and manual deletion.

⭐⭐ [Product Update] ChatGPT search now uses the site:operator at scale

Simon Willison cites Promptwatch tracking data showing that the share of ChatGPT Search fanout queries containing the site: operator jumped from about 0.3%-0.5% to 16%-17% on August 8 around the GPT-5.6 rollout, with a brief dip on August 3-5 suggesting staged deployment. The change aligns with OpenAI's August 6 announcement about updating GPT-5.6 Sol for better factual reliability. Simon speculates the latest search tool likely takes a form like search(query, recency, domains) rather than directly prompting the site: syntax. Promptwatch also reported on August 18 that ChatGPT appeared to greatly reduce Reddit usage in searches, though Simon could not find corresponding changes in leaked system prompts.

⭐⭐ [Product Update] The /wayfinder Skill: Navigating the “Fog of War” of Planning

Latent Space interviews Matt Pocock about his new skill '/wayfinder', designed for planning greenfield projects or situations where the end state isn't clear — navigating what Pocock calls the 'fog of war'. The skill adds an orchestrator layer that takes over planning sessions, splitting work into multiple threads, doing prototyping and research, then pulling results together, so the user isn't constrained by context-window management during planning. The core design relies on two precise document types: a 'map' (all decisions made so far) and a 'ticket' (the specific task for a child session). Pocock's 'AI Skills for Real Engineers' project has over 220,000 GitHub stars and his YouTube channel has 347,000 subscribers.

Research

⭐⭐ [Research] A shot-scraper-style JSON API on Bun 1.4's new Bun.WebView

Simon Willison demonstrates a JSON API service built on Bun 1.4's new Bun.WebView. The roughly 150-line, zero-dependency TypeScript service can execute JavaScript and capture PNG/JPEG/WebP screenshots without Puppeteer or Playwright, exposing /javascript, /screenshot, and /healthz endpoints. Bun 1.4 is the first stable release since the Rust rewrite, claiming 2,900+ bug fixes and improvements in idle CPU, memory, and startup time. Willison measured via cgroups that running a full Chrome for complex pages needs about 192-256MB of RAM.

Other

⭐⭐⭐ [Other] [AINews] Poolside gets $12B reverse-execuhire to NVIDIA; founders stay for $1B, employees go for $6B, Infraco scaling to 7GW neocloud

According to a letter obtained by Newcomer, AI model startup Poolside struck a $6 billion non-exclusive licensing deal with Nvidia, which also invested $1 billion in Poolside at a $12 billion pre-money valuation and is hiring about 109 of its employees. Poolside's founders say this is neither an acquisition nor an acquihire, but a "reverse execuhire": the founders and part of the team remain, while employees and investors get major payouts. Poolside previously lost a 40,000-GB300 cluster after failing to raise $2 billion in a six-week window. Its spun-out infrastructure arm, Poolside Infrastructure Company, is still building a 1.2GW datacenter in Texas and reportedly scaling toward a 7GW neocloud, and recently got a new CEO and CFO. The founders say they are not yet ready to share the updated vision.