Samwise Tech/AI/Robotics Newsletter
Thursday, June 25, 2026
OpenAI Unveils Jalapeño, Its First Custom AI Inference Chip
OpenAI unveiled Jalapeño, its first custom-built inference processor, developed in partnership with Broadcom. Designed specifically for running pre-built AI models at scale, Jalapeño delivers significantly better performance-per-watt than current alternatives, according to the company. OpenAI’s own AI models assisted in the chip’s design — a recursive development process the company says deepened its hardware instincts. President Greg Brockman described the chip’s origins: “We have a deep understanding of the workload. We’ve really been looking for specific workloads that are underserved.” The announcement signals OpenAI’s ambition to own its full infrastructure stack, from model development through chip architecture, memory systems, and deployment.
Sources: TechCrunch Share ↗ ✉︎ Email 💬 Text
Agility Robotics Plans $2.5B SPAC IPO for Its Digit Humanoid Robot
Agility Robotics announced plans to go public via a merger with Churchill Capital Corp XI, valuing the company at $2.5 billion. The deal is expected to generate more than $620 million in proceeds, including roughly $200 million from a PIPE financing backed by Amazon, Nvidia, and SoftBank Vision Fund 2. The Oregon State spinout, best known for its bipedal Digit robot, currently operates across nine customer sites including Toyota Motor Manufacturing Canada, GXO, Schaeffler, and Mercado Libre. CEO Peggy Johnson said “humanoid robots are poised to become a critical driver of productivity, supply chain resilience, and American technology leadership.” Agility plans to trade under ticker AGLT.
Sources: TechCrunch Share ↗ ✉︎ Email 💬 Text
Mistral Launches OCR 4, a 170-Language Document AI With On-Premises Deployment
Mistral launched OCR 4, a document intelligence model supporting 170 languages and every major document format — PDF, DOC, PPT, and OpenDocument — in a single deployable container. Priced at $4 per 1,000 pages ($2 via batch API), OCR 4 scored 85.20 on OlmOCRBench and achieved a 72% average win rate against competitors across more than 600 real-world documents. The release arrives as enterprise clients reassess their AI vendor risk following the June 12 export restriction that suspended access to Anthropic’s Fable 5 and Mythos 5. Mistral CEO Arthur Mensch has made AI sovereignty a central theme, and OCR 4’s on-premises deployment option directly addresses that demand.
Sources: VentureBeat Share ↗ ✉︎ Email 💬 Text
Another Wave of Top AI Researchers Is Leaving Google for Anthropic and OpenAI
Top AI researchers continue to exit Google for rivals. Researchers Jonas Adler and Alexander Pritzel, who played key roles in developing Google’s Gemini model, are joining Anthropic. The departures follow researcher Noam Shazeer — a Google veteran of over two decades — leaving for OpenAI last week, and Google DeepMind director John Jumper, who shared the 2024 Nobel Prize in Chemistry for his AlphaFold work, also departing for Anthropic. As OpenAI and Anthropic prepare for public offerings, the lure of pre-IPO equity makes recruiting top talent easier than ever. The sustained exodus raises questions about Google’s ability to retain the research depth that built its AI lead.
Sources: TechCrunch Share ↗ ✉︎ Email 💬 Text
Xiaomi’s HarnessX Rewrites Its Own AI Scaffolding Mid-Task — Smaller Models Gain the Most
Xiaomi researchers introduced HarnessX, a framework that treats an AI agent’s scaffolding as code to be evolved rather than fixed. A meta-agent powered by Claude Opus 4.6 analyzes logs and rewrites the harness mid-task, then tests the result on worker models including Claude Sonnet 4.6, GPT-5.4, and Qwen3.5-9B. Across 15 model-benchmark combinations, HarnessX delivered an average +14.5% performance gain — with Qwen3.5-9B seeing +44% on embodied planning tasks and +18.2% on SWE-bench. When harness-evolved data was used to fine-tune open-weight models, researchers observed an additional +4.7% boost. The findings suggest that improving the environment around a model can rival improving the model itself.
Sources: VentureBeat Share ↗ ✉︎ Email 💬 Text
Figma Adds Code Layers and AI Agent Skills to Its Collaborative Design Canvas
Figma’s annual Config update brings code layers directly to its collaborative canvas, letting teams clone repositories and extract flows from existing code into design layers for rapid prototyping. Chief Product Officer Yuhki Yamashita described the feature as designed for fast exploration rather than production-quality code, with designers, engineers, and product managers iterating in a shared spatial environment. The update also adds native support for animations, transitions, and 3D transforms, plus AI-generated shader effects. Users can now write natural-language prompts to build reusable AI agent skills within Figma, connect tools including Notion, GitHub, and Excel, and auto-generate custom plugins like layout generators or vector path tracers.
Sources: TechCrunch Share ↗ ✉︎ Email 💬 Text
Former Infosys CEO Vishal Sikka Raises $32M Seed for AI-Driven IT Services Challenger
Vishal Sikka, the former CEO of Infosys who spent twelve years building enterprise software at SAP, has launched Hang Ten Systems with a $32 million seed round led by Mayfield with strategic backing from Aramco Ventures. The company helps enterprises continuously build, modify, and operate software using AI-driven development and automation — a direct bet that AI will hollow out traditional IT services. Yahoo co-founder Jerry Yang sits on the board, and early customers include Siemens Gamesa Renewable Energy and Fresenius. The launch comes as Infosys shares have fallen more than 35% this year, and analysts at Jefferies forecast IT services among the first sectors to face meaningful AI disruption.
Sources: TechCrunch Share ↗ ✉︎ Email 💬 Text
Tech Pulse
Top Frontier Models: Claude Fable 5 (95.0% SWE-bench) | GPT-5.5 (88.7% SWE-bench) | Claude Opus 4.8 (88.6% SWE-bench)
Top Open Source Models: GLM-5.2 (91.2% GPQA) | DeepSeek V4 Pro (87.0% BenchLM) | Qwen3.5-397B (77.0% BenchLM)
Top Small Models (15–50B): Qwen3-30B (72.9% MMLU) | Mistral Small 4 24B (72.4% MMLU) | Gemma 3 27B (68.1% MMLU)
Top Edge Models (0–15B): Phi-4 14B (84.8% MMLU) | Phi-4-mini 3.8B (78.2% MMLU) | Gemma 4 E4B (65.1% MMLU)
AI Leaders: NVIDIA $5.14T | Alphabet $4.80T | Microsoft $3.40T
Robotics Leaders: Intuitive Surgical $156B | FANUC $26B | ABB Robotics $5.4B (private est.)
Curated by JD · samwise.agency
Leave a Reply
You must be logged in to post a comment.