# SkillOpt’s Agent Gains, Huawei’s Tau-Scaling Push, and Google’s App-Building Surge

*By AI High Signal Digest • May 26, 2026*

Microsoft showed that optimizing external skill files can sharply improve agents without retraining the base model, while Huawei outlined a packaging- and timing-centric chip roadmap and Google AI Studio’s Android builder drew mass early use. Also in the brief: small-model progress, new developer tools, DeepSeek’s funding push, and fresh legal pressure on AI training data.

## Top Stories

*Why it matters: the biggest developments today point to three leverage points in AI: better agent scaffolding, alternative chip-scaling paths, and faster consumer app creation.*

- **Microsoft Research’s SkillOpt showed large agent gains without changing model weights.** It treats the skill document as trainable external state for a frozen agent, with an optimizer model making validation-gated add/delete/replace edits. Microsoft said it was best or tied on all 52 tested model/benchmark/harness cells; on GPT-5.5 it added 23.5 points in direct chat, 24.8 with Codex, and 19.1 with Claude Code, with zero extra inference-time cost and transfer across models and harnesses [^1]. Another summary reported spreadsheet solving rose from 41.8% to 80.7% [^2].

- **Huawei used IEEE ISCAS to argue for a new semiconductor metric: τ-scaling.** The framework shifts focus from transistor geometry to time-based optimization across devices, chips, and systems [^3][^4]. In its paper, Huawei says LogicFolding on Kirin 2026 can raise density from 155 to 238 MTr/mm², improve energy efficiency by 41%, and increase frequency by 13%, with a roadmap to 400+ MTr/mm² and "equivalent 1.4nm" density by 2031; Kirin chips using the new architecture are slated to ship this fall [^4][^3].

- **Google AI Studio’s Android builder is scaling beyond developers.** Google said users created more than 250,000 native Android apps in the first week after launch, and likely over 99% of creators had never built an Android app before [^5]. That is a strong signal that prompt-driven software creation is already finding mass-market demand.

## Research & Innovation

*Why it matters: outside the headline stories, the most important research updates were about making models smaller, faster, and better at long-context memory.*

- **Gated DeltaNet-2 improved linear attention by separating erase and write operations.** The 1.3B model reportedly beat Mamba-3 and KDA head-to-head on language, reasoning, and retrieval, with S-NIAH-3 rising from 63 to 90 [^6][^7].

- **MiniCPM5-1B pushed the small-model frontier forward.** OpenBMB called it the strongest open-source base model under 2B parameters; it ranked #1 on Artificial Analysis’ small-model index at 17.9, ahead of Qwen3.5-2B at 16.3, and its ~0.5GB INT4 weights are designed for fully offline use on phones, browsers, and laptops [^8].

## Products & Launches

*Why it matters: launches were concentrated around developer tooling, cheaper inference, and image-generation workflows.*

- **xAI launched Grok Build, a coding CLI powered by Grok 4.3 Heavy, with a 2M-token context window and 8 parallel subagents [^9].**

- **Alibaba added automatic implicit caching to Qwen3.7-Max.** The feature activates with no setup and is positioned as faster and cheaper out of the box; users who need more deterministic hit rates can choose explicit caching [^10].

- **NVIDIA released PiD, a super-resolution model that works directly from model latents to deliver 4x resolution for generated images, with support for FLUX.1, FLUX.2, and Z-Image [^11].**

## Industry Moves

*Why it matters: the business signal is splitting between aggressive commercialization and harder questions about enterprise ROI.*

- **DeepSeek is reportedly seeking roughly $7.35B in fresh funding as rising compute costs push the lab toward commercialization [^12].** Separately, another report said DeepSeek cut model prices by 75% [^13].

- **Anthropic moved ahead of OpenAI in Ramp’s latest business-adoption index, 34.4% to 32.3%, but cost pressure is rising.** The same discussion pointed to image-inclusive prompts becoming 3x more expensive and said the fastest-growing vendors on Ramp are inference platforms selling cheap open-source models [^14]. In a separate enterprise datapoint, Uber said the link between AI consumption and shipped features is "not there yet" after burning through its 2026 Claude Code budget in four months, while slowing hiring to fund AI spend [^15].

## Policy & Regulation

*Why it matters: legal pressure is widening from companies to individuals, while governments are becoming more explicit about AI sovereignty.*

- **Two authors sued individual researchers over training-data practices, alleging Guillaume Lample torrented 70TB of pirated books to help train Llama and naming former Meta AI executive Joelle Pineau as involved [^16].**

- **Japan’s prime minister included domestic AI in a roundtable on the country’s "New Technology Nation" strategy.** Sakana AI said it discussed industry-specific AI deployments plus ways to use overseas models while preserving Japan’s defense autonomy and data sovereignty through domestic technology; the prime minister described domestic AI as one of 17 strategic fields where startups are opening new paths [^17][^18].

## Quick Takes

*Why it matters: a few smaller updates sharpened the picture on frontier reasoning, ethics, and robotics.*

- DeepMind follow-up posts said Gemini plus agentic loops has now solved 11 major open math problems, while Demis Hassabis said today’s systems are still "nowhere near" AGI [^19][^20].
- Pope XIV said the Church and Anthropic will work together to "find the way for humanity" in the age of AI [^21].
- LimX Dynamics opened global pre-orders for Luna, a 160cm commercial humanoid priced at RMB 298,000 in China, with claimed support for 200-unit fleet synchronization [^22].
- A deployment-aware context-optimization paper reported roughly 25% token savings at equal F1 and more than 50% lower token cost in high-performance settings on 5,000 HotpotQA instances [^23].

---

### Sources

[^1]: [𝕏 post by @omarsar0](https://x.com/omarsar0/status/2058936160291004483)
[^2]: [𝕏 post by @TheTuringPost](https://x.com/TheTuringPost/status/2059137971543269860)
[^3]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2058835932975915191)
[^4]: [𝕏 post by @ZhihuFrontier](https://x.com/ZhihuFrontier/status/2059118295580852374)
[^5]: [𝕏 post by @OfficialLoganK](https://x.com/OfficialLoganK/status/2058997700218294564)
[^6]: [𝕏 post by @ahatamiz1](https://x.com/ahatamiz1/status/2057586630450610673)
[^7]: [𝕏 post by @TheAITimeline](https://x.com/TheAITimeline/status/2059127219549606164)
[^8]: [𝕏 post by @OpenBMB](https://x.com/OpenBMB/status/2058903449379717319)
[^9]: [𝕏 post by @dl_weekly](https://x.com/dl_weekly/status/2058896246757113961)
[^10]: [𝕏 post by @Alibaba_Qwen](https://x.com/Alibaba_Qwen/status/2058932656797368619)
[^11]: [𝕏 post by @multimodalart](https://x.com/multimodalart/status/2059003125768339649)
[^12]: [𝕏 post by @theinformation](https://x.com/theinformation/status/2059016688645935246)
[^13]: [𝕏 post by @bfmbusiness](https://x.com/bfmbusiness/status/2058981358811185184)
[^14]: [𝕏 post by @kimmonismus](https://x.com/kimmonismus/status/2058893123414319441)
[^15]: [𝕏 post by @HedgieMarkets](https://x.com/HedgieMarkets/status/2059037292766146571)
[^16]: [𝕏 post by @ednewtonrex](https://x.com/ednewtonrex/status/2058433725889716519)
[^17]: [𝕏 post by @SakanaAILabs](https://x.com/SakanaAILabs/status/2058885261703201145)
[^18]: [𝕏 post by @takaichi_sanae](https://x.com/takaichi_sanae/status/2058858773876498575)
[^19]: [𝕏 post by @LeiYu63](https://x.com/LeiYu63/status/2059019777394237570)
[^20]: [𝕏 post by @ValerioCapraro](https://x.com/ValerioCapraro/status/2058870879807852659)
[^21]: [𝕏 post by @disclosetv](https://x.com/disclosetv/status/2058867784969343465)
[^22]: [𝕏 post by @humanoidsdaily](https://x.com/humanoidsdaily/status/2059040644002116051)
[^23]: [𝕏 post by @dair_ai](https://x.com/dair_ai/status/2058948732658626789)