Awesome Agents

More than 1,270 employees at the companies building the world's most capable AI models spent this week asking the U.S. government to develop the option of slowing them down. The statement, titled "Pacing the Frontier," went up on July 28 and has been collecting signatures ever since. When we checked the tally directly on the statement's own site, it stood at 1,273. The list of names is the part that makes this different from the usual open letter.

PJM Interconnection, the grid operator that keeps the lights on for 67 million people across 13 states and Washington, DC, told its members on July 27 that new AI data centers of 50 megawatts or larger won't get an unconditional promise of power anymore. Starting June 1, 2027, any facility that size without its own generation or a contracted supply faces curtailment the moment the grid runs short, ahead of the demand-response programs PJM already leans on during heat waves and cold snaps.

TL;DR Claude Opus 5 leads SWE-bench Verified among generally available models at 96.0%, priced at $5/$25 per million tokens - half of what Anthropic's own Mythos-class models charge for a similar score GPT-5.6 Sol reportedly matches it at 96.2%, but OpenAI hasn't published an official number and access is still gated to roughly 20 partners GLM-5.2 is the open-weight standout: 62.1% on SWE-bench Pro under an MIT license, the highest Pro score outside the Claude and OpenAI tier Three months...

Three papers landed this week that all circle the same uncomfortable question: how much do we actually know about what a model is doing when it isn't being watched, or when it's talking to someone in a language, or a mood, we didn't test for? One paper strips away the standard justification for alignment faking and finds it doesn't hold up. Another shows a single open-weight model gets meaningfully more deceptive depending on what language you speak to it in.

Qwen3-30B-A3B is the mixture-of-experts entry in Alibaba's Qwen 3 family, built to prove a point: a model that only turns on 3.3 billion of its 30.5 billion parameters per token can still compete with dense models several times its active size. Released April 29, 2025 alongside the rest of the Qwen3 line, it shipped with a hybrid thinking mode that lets a single checkpoint switch between step-by-step reasoning and fast direct answers depending on the prompt.

Every AI content detector on the market advertises a number in the high 90s. GPTZero's homepage says 99%. Copyleaks says 99%, "backed by independent third-party studies." Winston AI goes further and claims 99.98%, calling itself "the only AI detector" with that figure. Sapling is the outlier - it says 97%+ and then adds a caveat none of the others bother with: no detector should be used as a standalone check.

Three years ago, Core Scientific's data centers ran Bitcoin miners. On July 28, the company signed a 15-year lease putting more than 500 megawatts of that same footprint to work hosting AMD's GPUs instead, with an option to scale to 2.5 gigawatts. The deal, announced with Core Scientific's Q2 earnings, is worth more than $14 billion in base contracted revenue over its initial term.

Overview Muse Spark 1.1 is Meta Superintelligence Labs' second model release, and the first to ship with a public API. Meta launched it on July 9, 2026, exactly three months after Muse Spark debuted as the lab's first proprietary frontier model. Where the original Spark was a consumer-only release with no programmatic access, 1.1 flips that: it targets agentic coding, computer use, and tool orchestration, and it's available to developers through the new Meta Model API in public preview.

TL;DR 7.6/10 - the cheapest capable agentic model on the market, undercut by a geofence Meta hasn't explained Went from no public API and a weak coding score to a real, priced offering in three months flat Trails Claude Opus 4.8 and GPT 5.5 on every head-to-head coding benchmark, sometimes by a wide margin Worth testing if you're a US developer tuning for cost per agentic task; skip it everywhere else, because you legally can't use it yet Three months ago I wrote off Muse Spark as a consumer...

Cyera has agreed to acquire Oasis Security for approximately $1 billion, according to TechCrunch and confirmed by multiple Israeli outlets on July 28. The deal, split mostly cash with the rest in Cyera stock, is the fifth acquisition Cyera has closed this year and its largest by a wide margin. The target is a three-year-old startup that does one specific thing: keep track of what AI agents are allowed to touch.