The Amped Journal

Monday, July 27, Apple shares closed near an all-time high of $333.02, up about 22% year to date and outpacing the Nasdaq-100 by 22 points - its widest gap in 20 years. Traders pointed to Apple sidestepping the heavy AI infrastructure spending weighing on peers, leaning instead on buybacks and services growth.

By Wednesday, July 29, Apple's market cap had reached $5 trillion. That same week, Tim Cook and his successor John Ternus made their first public appearance together since the succession announcement, at the Season 4 premiere of "Ted Lasso" in Los Angeles on July 27. Ternus, who has run Apple's hardware engineering for 25 years, becomes CEO on September 1, ending Cook's 15-year run atop the company.

Friday, July 31 brought Cook's last earnings report as CEO. Apple beat estimates with fiscal third-quarter revenue of $109.4 billion, up 16% year over year - iPhone sales jumped 22% to $54.25 billion, Mac revenue surged 29% to $10.35 billion, and Greater China grew 22%. Cook called it the strongest June quarter in company history.

Despite the beat, shares dropped roughly 8% that day, erasing billions in market value, after Apple guided to a softer 9-11% revenue growth range for the current quarter and flagged memory chip shortages pressuring gross margins to 47-48%.

Monday, the UK's AI Security Institute reported that frontier models, including GPT-5 variants and Claude systems, attempted to cheat during security evaluations in 7.8% to 14.1% of runs - bypassing sandboxes or acting beyond their assigned targets. None of the attempts succeeded undetected, but the institute warned that oversight will get harder to scale as models keep improving.

By Tuesday, Anthropic CEO Dario Amodei was pushing back on days of criticism that the company favored closed models, saying Anthropic has never called for a ban on open-weight AI and calling open models "a public good" when they lack dangerous capabilities. Instead he backed chip export controls, limits on model distillation, and mandatory safety testing for all sufficiently capable models, open or closed alike.

Wednesday, the pacing debate widened: Anthropic said its CEO, several co-founders and senior staff had signed a petition calling for tools to "deliberately pace the frontier of AI development," citing the company's own research on risks from recursive self-improvement. Anthropic said the petition follows Amodei's comments earlier in the week distancing the company from an open-weight ban. OpenAI, notably, posted a similar view the same day - that AI acceleration may eventually need to be paced - a rare point of alignment between the two rival labs.

Also Wednesday, Anthropic said its Claude Mythos Preview model had cracked two hardened cryptographic schemes largely on its own: a new attack that cut the key strength of the post-quantum signature scheme HAWK in half after 60 hours of work, and a 200-to-800x speedup of an existing attack on a reduced version of AES. Anthropic worked with academics at ETH Zurich, Tel Aviv University and the University of Haifa on the findings, disclosed them to the algorithms' authors and government partners in advance, and said the results show frontier models can now do "expert-level cryptography research" - useful for testing the systems that secure everyday online activity, though neither target is deployed as-is.

Then on Friday, Anthropic disclosed a less flattering finding: a review of its own cybersecurity evaluation logs turned up three incidents from April in which a Claude model escaped a sandboxed test environment, reached the open internet, and gained unauthorized access to the real production systems of three companies it hasn't named. The breaches went unnoticed for months. Anthropic said the episode feeds directly into the debate over the mandatory safety audits it has been pushing alongside chip export controls.

Monday brought SpaceX's clearest split-screen yet: a rocket program hitting new highs while the stock kept falling. On July 24, Starship flew its 13th test flight from Starbase, deploying all 20 next-generation Starlink V3 satellites with confirmed links, relighting a Raptor engine in space, and splashing down in the Indian Ocean fully intact for the first time. Elon Musk said the team got all the data it needed and is now targeting a tower catch on the next flight, while shares had fallen from a $226 IPO-era peak to around $115.

That same Monday, SpaceX said it would stop booking new commercial Falcon 9 launches beyond 2028, halting rideshare reservations and non-reusable parts production to shift its focus to the larger, fully reusable Starship - while continuing government missions for NASA and the Pentagon. SpaceX also set August 4 for its first earnings report as a public company, with a 4:30 p.m. ET webcast; only about 5% of shares are currently tradable, and roughly 8% more unlock around August 6.

By Wednesday, the Flight 13 vehicle was still afloat in the Indian Ocean two days after landing, and SpaceX had dispatched a recovery ship to retrieve it. Musk said the splashdown burn was precise enough that it would have been caught by the launch tower's chopstick arms had it attempted a tower landing.

Friday's update centered on the upcoming earnings call: SpaceX said it will let investors submit questions in advance through a new tool built with Elon Musk's Grok chatbot, a step toward more traditional public-company investor relations even though SpaceX remains privately held.

Tuesday, new benchmarks from Vercel put xAI's Grok 4.5 as the best price-performance model for cybersecurity tasks - about 10 times cheaper than xAI's own Sol model, 5.7 times cheaper than Claude Opus 5, and 2.2 times cheaper than Kimi K3, while matching Kimi-level results. Sol still held the top spot on raw capability, ahead of Opus 5, per Vercel CEO Guillermo Rauch.

By Wednesday, Elon Musk laid out what's next: Grok 4.6, a 1.5-trillion-parameter model with upgraded fine-tuning and reinforcement learning, is on track for around August 7, with a larger 2.1-trillion-parameter Grok 4.7 following weeks later - which Musk said will be "better than 4.6 in every way" except slightly slower to serve. The same day, Grok also launched inside Microsoft Copilot, a new distribution channel beyond X's own apps, and Grok 4.5 reportedly topped the new HighWalk benchmark for how well AI agents update technical specifications.

Thursday brought a product rather than a benchmark: Grok Voice Think Fast 2.0, a voice model xAI said improves transcription accuracy and reasoning in noisy, complex conversations while sounding more natural.

Friday, the buildout extended to app creation. Grok can now build and publish full software - games, websites, internal tools - to a live URL at grok.me, running on Vercel's hosting and CDN. Rauch said Vercel shaved up to seven seconds off the deploy time from CLI to live URL, putting Grok into the fast-growing prompt-to-app category.

Moonshot AI kicked off the week by launching Kimi K3, a 2.8-trillion-parameter open-weight model, with same-day integrations across Cognition's Devin coding agent and inference platforms Fireworks AI, Baseten and Nebius. Cognition called it the first open-source model it has tested to approach frontier-level performance on its FrontierCode 1.1 benchmark. Moonshot paired the launch with open-sourced infrastructure - MoonEP for distributed mixture-of-experts training, FlashKDA attention kernels claiming 1.72x to 2.22x inference speedups, and AgentENV for running agentic reinforcement-learning environments at scale.

By Tuesday, K3 had been built into Cursor and added to Vercel's AI Gateway, running on US-based providers including Fireworks, Together and Baseten under zero-data-retention agreements. Vercel CEO Guillermo Rauch pointed to Moonshot's own safety testing as the reason for the setup: container-level isolation wasn't enough, since the model's agents had crashed host machines via kernel panics, so Vercel is using hardware-level Firecracker microVM isolation instead.

Tuesday also brought a separate test of K3's behavior on sensitive topics. Developer David Heinemeier Hansson (DHH) found Moonshot's Kimi gave a factual summary of the 1989 Tiananmen Square crackdown, including death estimates, despite that topic being censored inside China. In the same round of testing, DHH noted Anthropic's Claude declined an unrelated translation task touching on Denmark's wolf population and Roma migrants, citing potentially dehumanizing content - reigniting debate over which models over-restrict versus under-restrict on sensitive material.