curated · 2026-06-25

This report pulls together the big numbers on the AI economy. $175B annualized and the acceleration from 180 days to under 2 days to add $1B in revenue—demand is compounding fast. A 10% token price drop driving 12–18% usage growth shows real price elasticity, so the scale-via-price-cuts playbook is working for now. The supply side is less rosy. Hyperscaler AI revenue barely covers infrastructure depreciation, and GPU economics rest on a six-year useful-life assumption. Power and data center costs remain the hard physical ceiling. I'd read this as a solid demand snapshot; the profit story needs next quarter's earnings to back it up.

I clicked because this is the first time a US administration has directly blocked an OpenAI frontier model release. The Verge's report is thin: the Trump administration asked for a delay, and future customer access will be approved case-by-case. No word on how long, what the approval criteria are, or whether OpenAI pushed back internally. I'd read this as administrative pressure, not a formal safety review — no agency name, no evaluation process, sounds more like a phone call from the White House. If Commerce or NIST issues paperwork later, that changes the story. For now, too many blanks to call it an export-control escalation.

Codex was desktop-only until now, so getting a mobile remote is a real workflow unlock. OpenAI added 1:1 device pairing—your phone binds to one specific machine, which is tighter than a random browser login. The key thing: the phone doesn't run Codex. Your laptop or Mac mini does the heavy lifting in the background. The app handles notifications, goals, side chat, file previews, and inline review comments. If you already use Codex, this means you can check progress and approve next steps without being at your desk. If you don't, the mobile app isn't a standalone reason to start—it's an extension of the desktop experience. The post doesn't say whether Android and iOS launched together, and I haven't seen details on notification latency or background reliability yet. Those are the two things I'd watch before getting excited.

This is worth opening because it's the first time the US government has directly blocked an OpenAI model release. GPT-5.6 was supposed to go wide; now it's a controlled preview for a small set of partners, with the government approving each customer. Sam Altman confirmed the process to staff on Thursday, so this isn't a rumor. The concern is specific: the model can automate high-skill cyber work. Defenders finding bugs faster is good, but attackers speeding up exploit testing is the obvious flip side. That dual-use capability is what triggered the intervention. Right now we only have The Information's summary—no word on which agency made the call, what the approval criteria are, or how many partners get access. If someone from AISI or Commerce speaks up, this story gets heavier.

This one's worth opening because the WaPo methodology is cleaner than most political-bias studies: they didn't ask models to self-identify, they checked whether answers only presented one side's arguments. GPT-5.5 gave left-only answers 80% of the time, DeepSeek V4 Pro 70%. Both opposed the death penalty, even though Gallup shows majority US support for decades. The more interesting bit is the two "anti-woke" models. Grok 4.3 did produce more right-leaning answers than any other model tested, but still skewed left more often. Gab's Arya, explicitly marketed as conservative, leaned left even more. The one outlier: Gemini 3.1 Pro presented both sides in 93% of cases. I'd discount this a bit — WaPo didn't release the full prompt set or explain who labeled answers left/right. But the article is explicit that Grok took a fully right-leaning stance on trans rights, matching Musk's public position. That tells me xAI can steer when it wants to; it just didn't do it across the board.

Google shipped Computer Use in Gemini 3.5 Flash—the model can now see screens, click buttons, scroll, and run long tasks across browsers and phones. The useful bits: built-in mobile and desktop OS support so you don't scaffold it yourself, intent arguments on every function call to help you figure out what the model wants to do, and customizable human-in-the-loop handoffs for critical moments. The post doesn't disclose pricing or latency, which are the two things that make or break this in production. Computer Use pipelines have long call chains and heavy inference overhead, and Flash is positioned as a lightweight model—so real-world speed and cost are total unknowns. I'd treat this as an experimental tool for automated QA or fixed workflows, not something to push to prod yet.

The logic isn't crazy: keystrokes, mouse movements, and decision sequences in games carry timing and physical feedback that pure text doesn't. Pim de Witte wants to train models on millions of hours of this and sell them to robotics firms and game devs. But the post leaves out two things I'd want before getting excited: valuation and investor names. Without valuation, we don't know what price the market put on this narrative. Without backers, we can't see who's vouching for it. I'd treat this as a large-scale experiment in gameplay pretraining. Interesting direction, but "general physical intuition" still has to clear validation, generalization, and real robot deployment — none of which are in the article.

This is worth a click because it squeezes a few extra percentage points of throughput out of DeepSeek V3/R1/V4 inference with no accuracy trade-off. Waterfill is the simpler one: it sends the shared expert to whichever GPU is least busy right now. On DeepSeek V4, that takes throughput from 49,253 tok/s to 51,677 tok/s. LPLB uses linear programming to route tokens across redundant expert replicas, adding up to 7.34% on V3/R1. Both are already merged into SGLang main. I'd discount this a bit: the post reports throughput but stays silent on latency. For online serving, latency matters more, and we don't know if these methods add tail latency. The tests also ran on two Hopper nodes—gains at larger scale are unproven. If you're already serving DeepSeek models on SGLang, these flags are worth toggling on, but don't expect them to fix all load imbalance. They handle residual dispatch-time skew, not the underlying expert placement problem.

This one's worth opening because it pushes agentic coding scores to a new open-source high: 82.4 on SWE-Bench Verified and 77.5 on Terminal-Bench 2.1, both best-in-class for open models right now. The lineup runs from 9B to 397B MoE, all MIT-licensed with GGUF versions ready, so the local-run barrier is low. The training approach is the interesting bit—they used RL to jointly optimize the task scaffold and the final solution, letting the model improve its own execution framework rather than just memorizing answers. Base models are post-trained gemma4 and qwen3.5, not trained from scratch. I'd discount this a bit: the post doesn't disclose training cost, inference latency, or hardware requirements. A 397B MoE won't be light to run locally even with GGUF quantization. SWE-Bench scores have also inflated over time, and there's no comparison showing where 82.4 lands relative to closed-source models. Treat it as a strong open-source, commercially-usable baseline for agentic coding—don't rush to benchmark it against Claude Sonnet 4.5.

This one's worth a click because it turns 'does the agent remember what you said last week' into a measurable test. VitaBench 2.0 simulates 56 virtual users, each with roughly 4 years of interaction history and 2,093 events—tasks cover remembering preferences, asking proactive questions, and calling 66 tools. Even Claude Opus 4.6, the top model, barely averaged above 0.5 in open-book mode. Two details stand out: thinking mode didn't consistently help on personalization tasks—sometimes overthinking drifted away from the user's actual intent—and all models saw a cliff-like drop on tasks requiring proactive questions. Models are still much better at passive answering than at clarifying ambiguity. I'd discount this a bit since it's simulated data, and the gap from real user behavior hasn't been validated. But as an engineering benchmark, it gives teams building memory layers and agent workflows something more business-relevant than needle-in-a-haystack tests.

The numbers and the internal pushback both check out. Meta already routed roughly half of moderation requests through LLMs in 2025 and wants to push that above 90% for some content types by year-end. The company claims 13% fewer errors and 10% more violation catches versus humans, saving billions a year. Employees say the models still remove or shadow-ban harmless content and that oversight hasn't kept up. Behind the scenes, Meta is swapping Google Gemini for its own Muse Spark model, trained on past human decisions. I'd discount the 13% and 10% figures a bit — those come from Meta's own tests, and we don't see the eval set or how false positives are counted. The part I'd actually watch is shadow-banning: silent throttling is hard for creators to detect and even harder to appeal, which makes it more prone to overuse than outright takedowns.

This caught my eye because it's Anthropic accusing Alibaba — two heavyweights. But the post is just a one-sentence summary: Anthropic claims Alibaba reverse-engineered Claude's core tech. No word on how, which parts, or what evidence they have. Alibaba hasn't responded. I'd discount this for now. Reverse-engineering closed models isn't new, but a legal claim at this level usually implies internal data leaks or API abuse, not just clever prompting. With only a headline, we can't tell if this is distillation or something more serious. Wait for actual filings.

This is worth a look because it fixes a real friction point: you're coding in Claude Code or Cursor, you ask your agent to pick the best model for a task, and it's working off training data from six months ago — wrong prices, stale benchmarks. OpenRouter packaged its model catalog, Artificial Analysis intelligence scores, Design Arena ELO rankings, and live pricing into a remote MCP server you can add with one command. The two most useful tools: filtering models by price and context length, and firing test prompts to compare responses and costs across models. First login mints a dedicated key with a 7-day expiry and a $10 spend cap, so you won't accidentally blow your budget. The post doesn't say whether the MCP server itself is free or paid — I'd flag that as an open question. Also, this feels more like a model-selection plugin for coding agents than a general-purpose agent toolbox. If you regularly switch between models, it'll save you a lot of tab-switching.