curated · 2026-06-24

The useful bit here is that computer use is now a native tool inside Gemini 3.5 Flash, not an external wrapper. That means fewer moving parts and likely lower latency. But the post doesn't share any benchmark scores or latency numbers—just that the API is live. I'd discount the hype until someone tests it on real-world tasks. Built-in tooling is cleaner architecturally, but reliability on messy UIs is still an open question.

This is worth a click because it's a concrete integration path, not a vague partnership announcement. Notion treated the Cursor SDK as the agent engine and built only the surface and context layer. A Notion thread maps to a Cursor agent, each message becomes an agent run streamed over SSE, and they hooked up their own remote MCP server so the agent reads and writes workspace state in real time. Engineer Victor Shen called the integration "a thin adapter" because the SDK's agent/run model lined up almost directly with Notion's threading model. I'd discount this a bit: it's a customer case study on Cursor's own blog, not an independent review. The post doesn't disclose launch date, pricing, or whether this is available to all Notion users or still internal. But the technical specifics—remote MCP, cloud sandboxing, auto PR creation—are real enough that if you're thinking about embedding coding agents into your own product, this is more useful than most "we added an API" stories.

This one's worth a look because Figma is in a weird spot. On the product side, they're adding everything: Code Layers puts design and production code side by side, Motion brings animation timelines into collaborative editing, Shader uses WebGPU for material effects. It looks like they want to lock designers into an all-in-one canvas. But the financial side tells a different story. Figma admits that third-party AI inference costs are squeezing margins. Worse, those models come from Anthropic and others who are building tools that generate interfaces directly. Figma's counter-bet is on human judgment — AI produces tweakable half-finished work rather than one-shot outputs, plus team-shared prompts and plugins to cut token use. I'd discount the strategy a bit: the post doesn't spell out progress on in-house models, just says "sort of." If core AI depends on competitors long-term, all the product features in the world won't fix the cost structure or the strategic leverage problem.

The reason to click: all three major labs are now shipping the same bet—let the model see the screen, click buttons, and fill forms. Google put computer use into Gemini 3.5 Flash, a model built for low latency and low cost. The positioning is clear: this isn't for complex desktop workflows, it's for high-volume narrow tasks like web form filling and data entry. The post claims Flash beats Claude Sonnet 4 and GPT-5 on WebVoyager, but doesn't release test splits, retry counts, or prompt templates. I'd discount the numbers until third-party repros land. Pricing and rate limits aren't disclosed either. Flash is usually cheap, but computer use means screenshots plus inference per step, so real token burn will be higher than plain text. The API and AI Studio are live—worth running a few simple tasks to check per-step cost before scaling.

This lands because of who's talking: Hoffman is an investor in both OpenAI and Anthropic, and a former Microsoft board member. His SpaceX read is concrete — post-IPO Cursor acquisition is using market cap to buy an AI label, and the compute leasing business is just 'a premium-priced CoreWeave.' On xAI, he's brutal: all 11 co-founders gone, company on its third restart. I'd discount this a bit — he's literally invested in the competitors. But the '11 co-founders departed' and 'third restart' facts are publicly documented, not his invention. His take on the government forcing Anthropic to take down Fable and Mythos models is also worth noting: he calls it 'autocratic willy-nilly,' and points out Anthropic flagged the security issues before the order came down. What's missing: any response from SpaceX or xAI. The article doesn't say if they pushed back.

This one's worth opening because OpenAI finally turned its custom chip rumors into actual engineering samples. Jalapeño, built with Broadcom, went from design to tape-out in 9 months—fast by chip industry standards—with OpenAI's own models helping accelerate the process. Engineering samples are already running GPT‑5.3‑Codex‑Spark at target frequency and power in the lab. OpenAI says this isn't a repurposed general accelerator; it was architected from scratch around the serving patterns of ChatGPT, Codex, and future agentic products. Early testing shows substantially better performance per watt than current state-of-the-art, but detailed benchmarks won't land for a few months. I'd discount the "substantially better" claim a bit—no comparison baselines or workloads are specified, so it's just OpenAI's word for now. The real thing to track is whether the gigawatt-scale deployment actually starts in 2026, and whether Microsoft is in the first wave. If this scales, OpenAI's negotiating position with NVIDIA improves dramatically.

The reason this is worth a look: Bidi 1 fixes the old pain point where you have to wait for the AI to finish speaking before you can jump in—nothing like a real conversation. Some users already see it in ChatGPT's web and app model selector; picking it turns the voice bubble yellow. A demo shows someone asking it to count from 1 to 10, interrupting mid-way with "count backwards," and it switches instantly. OpenAI hasn't announced a launch date. One outlet guesses a wider rollout this week. The post doesn't cover pricing, regional availability, or how well it holds context after interruptions. I'd treat this as a feature preview for now—latency and context handling after back-to-back interruptions are what I'd watch for once it ships.

The reason to click: Qwen trained an agent that predicts what the environment will look like before taking an action, then open-sourced both the model and the benchmark. They call it a "native language world model." The concrete part: it covers seven domains—MCP, search, terminal, software engineering, web, OS, and Android—trained on over 10 million real interaction trajectories through a CPT→SFT→RL pipeline. On their in-house AgentWorldBench, the 397B-total MoE model scored 58.71, just ahead of GPT-5.4 at 58.25 and Claude Opus 4.8. The margin is small, but the open-source angle matters more. You can use it as a decoupled environment simulator for Sim RL—on WideSearch it hit 50.3% F1, beating real-environment RL at 45.6%. Or use it as an agent foundation model: with LWM warm-up, it transfers to seven benchmarks, three of which never appeared in training. That transfer is more interesting than the raw score. Where I'd discount: the benchmark is their own, no external validation yet, and a 397B-parameter MoE isn't cheap to run. But the open release makes this worth tracking.

The reason to click: Doubao is packaging its agent-capable model as a paid office product. Free users get the Turbo model; Pro unlocks Doubao 2.1 Pro, which can control your local computer and browser, invoke Skills, schedule tasks, and generate online apps with a backend database. Pricing: Standard ¥68/month, Enhanced ¥200, Advanced ¥500. Verified students get Standard at ¥38/month for six months. I'd discount this a bit. The post lists features and prices but skips concurrency limits, task latency, and context window. No word on how reliably the agent controls a local machine either. Right now it looks like a paid entry point for office workflows—whether it's worth it depends on real-world stability and speed of the Pro model.