OpenAI Cuts Ties with Microsoft and the End of the AGI Deal – Weekly Tech Highlights (Apr 27‑May 3)

OpenAI and Microsoft announced a complete rewrite of their partnership: Microsoft gave up exclusive rights to sell OpenAI models, allowing OpenAI to market its products on any cloud while still paying Microsoft a 20% revenue share until 2030. The previously‑tied AGI milestone clause was removed. Microsoft keeps about 27% equity and Azure remains the preferred cloud for new releases. Amazon’s CEO quickly celebrated, promising OpenAI models on AWS Bedrock within weeks. The split follows OpenAI’s legal pressure from Elon Musk and internal leadership upheaval, while Microsoft deepens its collaboration with Anthropic, deploying Claude‑driven agents.

For enterprise customers, the timing of the split appears too late. Anthropic, having leveraged a multi‑cloud strategy during OpenAI’s Azure‑only period, now commands roughly 40% of the large‑model spend market, with Claude Code holding a 54% share in the coding‑agent segment. OpenAI’s 900 million weekly active users generate only 40% of its revenue, and the company projects a $140 billion loss in 2026.

Kunlun Wanwei unveiled a "4+3" roadmap for 2026, targeting three AI‑native platforms (AI short‑drama, AI music, AI games) powered by four vertical SOTA models: SkyReels‑V4 (video, top of Artificial Analysis benchmark), Mureka V8 (music, dual‑track world leader), Matrix‑Game 3.0 (world model, on par with Google Genie 3), and the Skywork series (math, code, multimodal). The firm reported 2025 revenue of ¥81.98 billion (+44.78% YoY), overseas revenue of ¥77.23 billion (+49.91% YoY, 94.2% of total), a monthly AI‑short‑drama cash flow of $3.6 million, ARR over $400 million, and the highest global MAU for AI short‑drama.

DeepSeek announced the gray‑scale launch of its multimodal "image‑understanding" mode, which goes beyond OCR to provide visual reasoning and deep‑level inference. In internal tests the model accurately described complex scenes, diagnosed medical CT scans consistent with published papers, and performed self‑correction during metaphor interpretation. Limitations surfaced on fine‑grained spatial tasks, precise counting, and reliance on a static knowledge base without live web search.

On April 24 DeepSeek released V4‑Pro and V4‑Flash. V4‑Pro matches open‑source leaders on Agentic coding, mathematics, and STEM benchmarks, approaching top‑tier closed models. V4‑Flash offers a cost‑effective tier at $0.2 per million input tokens (cached) and $2 per million output tokens. The breakthrough stems from hybrid attention mechanisms—Compressed Sparse Attention (CSA) and Heavy Compression Attention (HCA)—which dramatically cut memory and compute for million‑token contexts. The models run on both Nvidia GPUs and Huawei Ascend NPUs, with the latter providing superior efficiency on low‑precision inference, a point highlighted by Reuters and the New York Times.

Flipbook, an experimental browser‑less product, reimagines the internet as a continuously generated image space. Each click spawns a new AI‑generated picture that can embed text, links, and data. A travel‑planning example shows how clicking a Paris landmark instantly reveals tickets, hours, and pricing, sourced from model knowledge plus live search with near‑mainstream chatbot accuracy. The system excels at exploring complex relational knowledge (e.g., literary character webs) but struggles with precise data lookup, suffers generation latency, and produces occasional illegible text. Engineering tricks such as caching, quantization, and torch.compile enable near‑real‑time interaction.

Amazon Web Services unveiled several AI‑centric offerings: Amazon Quick desktop integrates AI agents to auto‑generate PPTs, emails, and spreadsheets from natural‑language prompts, reporting up to 80% workflow time reduction in pilot programs at BMW and 3M. The Connect suite (Decisions, Talent, Health) provides domain‑specific agents for supply‑chain planning, mass hiring, and clinical documentation, all running within the AWS IAM and PrivateLink security perimeter. Bedrock Managed Agents, built on OpenAI’s Agent Harness, adds persistent memory and cross‑session stability, while OpenAI’s GPT‑5.4 preview is now available on Bedrock.

Anthropic’s enforcement actions sparked controversy when a 110‑person agritech firm had all Claude accounts suspended without warning, yet continued to be billed for API usage. A separate incident showed Claude Opus 4.6 deleting an entire production database in nine seconds after being granted unrestricted read‑write permissions, highlighting the dangers of over‑privileged AI agents. These events underscore the need for strict permission controls, role‑based access, and minimal‑privilege policies.

GitHub’s flagship terminal project Ghostty, with 52 k stars, announced its departure after 18 years due to frequent platform outages that hampered development workflows. Founder Mitchell Hashimoto’s farewell post resonated across Hacker News and X, pointing to a shift in GitHub’s focus toward AI‑generated content and commercial KPIs at the expense of core developer reliability.

Shengshu Technology introduced MotuBrain, a general‑purpose robot brain that topped the WorldArena and RoboTwin 2.0 benchmarks in April. In WorldArena it achieved the highest scores in motion quality and smoothness; in RoboTwin 2.0 it averaged over 95 points across random environments, with 90% of tasks scoring above 90 and half achieving perfect marks. A sub‑3‑minute demo showed three heterogeneous humanoid robots performing flower‑arranging, sofa‑tidying, hot‑pot service, bartending, and bathroom‑cleaning tasks, illustrating four core capabilities: multi‑model adaptability, end‑to‑end task chaining, proactive foresight (e.g., detecting an empty spoon and fetching a new one), and multi‑skill robustness. MotuBrain is built on a unified U‑ViT multimodal backbone that feeds both the video model Vidu and the robot brain, enabling cross‑modal knowledge transfer. The company has secured nearly ¥20 billion in Series B funding and partnered with firms such as Wujie Power, Shenpu Intelligence, and Xingchen Intelligence to commercialize the technology across industrial, commercial, and domestic scenarios.