AI · Daily · June 25, 2026

chips, fast models, and the hidden frontier

OpenAI's Jalapeño chip is reported to push state-of-the-art perf per watt for LLM inference, with a design tuned for memory movement and networking, and likely wafer-scale packaging.[3]

Separately, OpenAI and Broadcom announced an LLM-optimized inference chip, reinforcing the shift toward custom silicon for model serving rather than generic GPUs alone.[4]

GLM‑5.2 FAST reports up to 250 tokens per second in serverless setups with a 1M-token context window, while remaining cost-competitive with closed models.[5]

Kog.ai's 2B model processes about 3,080 tokens per second, and SDXL can now run locally in browsers via WebGPU for offline image generation.[6]

Consumer NPUs look anemic by comparison, with AMD's Strix Halo NPU managing roughly 12 transactions per second on Gemma e4B and suffering memory-bandwidth and tooling issues, even as KV-cache and latent-space communication methods report substantially faster multi-agent inference and higher task performance.[7]

three blocs, three stories of 'frontier'

The EU is backing EUROPA, an open-source model with over 400 billion parameters to be trained on European supercomputers, pitched as affordable AI for public services.[8]

Qwen‑AgentWorld‑35B‑A3B scores 56.39 on AgentWorldBench and outperforms Claude Sonnet 4.6 on that benchmark.[9] GLM‑5.2 is reported to match Claude Opus on 45 coding-agent tasks while running at less than half of Opus's cost.[10]

DeepSeek's R1 model is highlighted for competitive reasoning ability at far lower training cost than US frontier models, and Kimi DeepSeek MiMo workflows are reported to rival pricier systems.[11]

US labs counter with hardware and branding power, from OpenAI's Jalapeño chip to GPT‑5.5 Instant, even as Amnesty International labels leading systems like DeepSeek and Gemini 'unlawful by design' and EU commentators doubt EUROPA's competitiveness.[12]

At the same time, Anthropic accuses Alibaba's Qwen lab of illicitly extracting Claude capabilities via around 29 million chats from roughly 25,000 fake accounts, while export controls and lawsuits over the banned Fable 5 model signal tightening control of advanced weights.[13]

agents, not models, are where the weirdness is

Qwen-AgentWorld simulates seven environments, including MCP and Android, using a 35B-parameter MoE model with about 3B active parameters per token.[14]

That model scores 56.39 on AgentWorldBench and outperforms Claude Sonnet 4.6 on that benchmark.[15] MCP tooling is exploding around this, with generators like mcpgen that turn OpenAPI specs into Python MCP servers, dynamic tool registries such as Heku, and policy enforcers like Imara tightening per-agent permissions.[16]

Runlayer just raised $30M to connect MCP-style agents to enterprise oversight systems, while many teams prefer self-hosted AI tools over cloud APIs for security and compliance.[17]

Grok Build v0.2.64 adds a dashboard for directory management and agent dispatch alongside an official MongoDB plugin, and Anthropic's Claude Tag feature plus LangChain/LangGraph stacks for airline-policy RAG show platforms converging on persistent, tool-rich coworker agents.[18]

coding copilots vs the craft backlash

Companies are standardizing on coding agents like AWS Kiro, GitHub Copilot, Cursor, Codex, and Claude Code, with GLM‑5.2 and Qwen 3.6 27B also used as agentic coders on commodity 32GB GPUs.[19]

A recent benchmark of eight AI coding agents found that none asked clarifying questions before planning and only one produced production-ready code, at a run cost of about $1.94.[20]

Developers report that reviewing AI-generated code is often harder than writing it, especially for juniors, and many teams disable AI tools during code review to avoid leaking proprietary client code to external APIs.[21]

Hacker News threads around Craft Capitallab describe an identity crisis as engineers see rising bug rates and burnout blamed on AI-written code, while major open-source maintainers explicitly reject AI-generated patches.[22]

Linus Torvalds publicly mocked claims that 99% of code is written by AI and attacked Linux sched_ext's design until its authors rewrote it, even as practitioners note the real bottleneck has shifted from typing to planning and system design.[23]

multimodal is splintering into weird little guilds

Anima, trained heavily on Danbooru-style datasets, is now the default for anime images on modest hardware, with its 2B-parameter model preserving diverse characters and styles more reliably than 12B-parameter Krea 2.[24]

Krea 2 and its Turbo variant shine on prompt adherence, text rendering, and style adaptation, running in GGUF form on roughly 8GB of VRAM and producing images in a few seconds on a 3090-class GPU, but struggle with anime, emotional expression, and backgrounds.[25]

A new realism LoRA for Krea 2 produces high-quality SFW and NSFW images, and character LoRAs are reported trainable on 16GB GPUs with 64GB RAM for under $3 via Runpod.[26]

Using a 320MB LoRA with Krea 2 Turbo, however, can stretch per-image times from seconds to minutes and often needs tricks like disabling prompt_enhance to keep results usable.[27]

On the video and document side, Seedance 2.5 can generate 30‑second 4K clips from a single prompt, HappyHorse 1.1 targets dynamic expressiveness, and Baidu's 3.3B-parameter Unlimited-OCR model under MIT license can parse dozens of pages in one pass.[28]

Users still rate Gemini's vision features and tools like a 2‑second plant disease scanner above Claude's vision, but complain about hallucinations and overbearing guardrails even as they lean on these models for search-like tasks.[29]