Samwise Tech/AI/Robotics Newsletter
Tuesday, June 23, 2026
Reflection AI Signs $6.3 Billion Compute Deal with SpaceX for Nvidia GB300 Access
Reflection AI, an open-source AI startup, signed a compute deal with SpaceX worth up to $6.3 billion through 2029, paying $150 million per month for access to Nvidia GB300 chips at SpaceX’s Colossus 2 data center. The agreement, confirmed June 22, gives Reflection priority capacity on the same facility serving Anthropic and Google. Either party may exit with 90 days’ notice after the first three months. SpaceX now reports more than $2.3 billion in monthly contracted AI compute revenue, marking one of the largest compute contracts ever signed by an open-source AI company.
Sources: TechCrunch
NVIDIA Launches Halos for Robotics, the Industry’s First Full-Stack Safety System for Physical AI
NVIDIA announced Halos for Robotics on June 22, calling it the industry’s first full-stack safety architecture for physical AI systems and humanoid robots. Drawing on more than 18,600 engineering years of autonomous vehicle safety work, Halos spans four layers: IGX Thor industrial compute hardware, Holoscan Sensor Bridge for sensor connectivity, Halos OS safety software, and an AI Systems Inspection Lab for certification preparation. Agility Robotics is the first company to integrate the platform into its Digit humanoid, already deployed at Amazon, GXO, Schaeffler, and Toyota. The launch addresses growing industry concern about safely deploying autonomous machines in shared human workspaces.
Sources: The Robot Report
Groq Confirms $650 Million Raise as AI Inference Cloud After Nvidia’s $20 Billion Technology Deal
AI chipmaker Groq confirmed a $650 million funding round on June 22, roughly six months after Nvidia signed a $20 billion agreement to license Groq’s chip technology and hired away founder and CEO Jonathan Ross along with president Sunny Madra. The new round, led by Disruptive and hedge fund Infinitum, positions Groq as an AI inference cloud provider after the leadership overhaul. The company, last valued at $6.9 billion in September, declined to disclose a new valuation. Groq is now rebuilding its executive team and leaning into its neocloud business, offering fast inference at scale without depending on its original hardware design team.
Sources: TechCrunch
Copilot SearchLeak and LiteLLM Admin Exploit Expose AI Trust Boundary Gap in Enterprise Stacks
Two separate AI security vulnerabilities exposed critical gaps in enterprise deployments in mid-June 2026. Varonis disclosed SearchLeak (CVE-2026-42824), a proof-of-concept attack in which a crafted Microsoft link triggers Copilot to search a victim’s mailbox and exfiltrate data through a Bing server-side request forgery. Separately, Obsidian Security revealed a three-CVE chain against LiteLLM, the AI gateway proxy used by thousands of enterprises, allowing privilege escalation to admin and remote code execution. LiteLLM CVE-2026-42271 appeared on the CISA Known Exploited Vulnerabilities catalog with a June 22 remediation deadline. VentureBeat noted both incidents broke the same AI trust boundary within a two-week period.
Sources: VentureBeat
Alibaba’s HappyHorse 1.1 Rises to Global No. 2 in AI Video as OpenAI Discontinues Sora
Alibaba Cloud released HappyHorse 1.1 with full API access and a 40 percent launch discount, as the AI video generation market reshuffled following two major withdrawals. OpenAI discontinued Sora after it proved financially unsustainable, and ByteDance indefinitely shelved the international rollout of Seedance 2.0 following copyright complaints from Hollywood studios. HappyHorse 1.1 now ranks second across major independent benchmarks, scoring 1,444 points in both text-to-video and image-to-video categories, 69 points ahead of Google’s Veo-3.1. Backed by $52.7 billion in Alibaba infrastructure investment, the model offers full enterprise API access and is positioned as a production-ready alternative for developers previously relying on Sora or Seedance.
Sources: VentureBeat
Self-Harness Framework Lets AI Agents Rewrite Their Own Operating Rules, Boosting Performance 60%
Researchers at the Shanghai Artificial Intelligence Laboratory introduced Self-Harness, a framework enabling large language model-based agents to systematically rewrite and improve their own operating rules during task execution. Unlike static rule sets, Self-Harness allows an agent to identify underperforming instructions, generate alternatives, and update its behavioral guidelines mid-run, boosting performance on evaluated tasks by up to 60 percent in testing. The approach targets a longstanding limitation of agentic systems where fixed prompts degrade as task conditions change. Researchers say the paradigm could substantially reduce the manual prompt engineering required to maintain high-performing AI agents in production environments, though the system remains in research settings for now.
Sources: VentureBeat
Claude Code Creator Boris Cherny Says AI Agent Loops Are as Big a Shift as the Move to Agentic AI
Boris Cherny, the creator of Claude Code, told attendees at Meta’s @Scale conference on June 22 that AI systems built around loops — where agents prompt other agents to generate and run code — represent a shift as significant as the transition from source code to agentic AI. Cherny described loops as architectures in which one model outputs a task, a second writes code to fulfill it, and the result feeds back into a new agent call. This compounding pattern accelerates development but introduces failure modes when loops degrade. Cherny said managing loop behavior, not just model capability, will define the next generation of AI tooling.
Sources: TechCrunch
Tech Pulse
Top Frontier Models (SWE-bench Verified): GPT-5.5 (88.7%) | Claude Opus 4.8 (88.6%) | Claude Fable 5 (95.0% — export-restricted)
Top Open Source Models (AI Index v4): Kimi K2.6 / Moonshot AI (53.9) | MMo-V2.5-Pro / MiniMax (53.8) | DeepSeek V4 Pro (51.5)
Top Small Models (15–50B): Qwen3-30B (est. top small tier) | MiniMax M3 (SWE-bench Pro 59.0%) | GLM-5.2 coding (SWE-bench Pro 62.1% at 753B)
Top Edge Models (0–15B): Phi-4-mini (strong reasoning, edge-optimized) | Gemma 3-9B (Google on-device) | Qwen2.5-7B (widely deployed)
AI Leaders (market cap): NVIDIA $5.4T | Alphabet $4.6T | Microsoft $3.1T
Robotics Leaders: Intuitive Surgical $146B | Figure AI $39B (private est.) | Neura Robotics $7B (private est.)
Curated by JD · samwise.agency
Leave a Reply
You must be logged in to post a comment.