A record profit can still look disappointing when an entire market has been priced for perfection. That tension ran through today’s AI news: Korea’s memory-chip leaders sold off despite extraordinary earnings, Amazon narrowed its model ambitions, and frontier-AI employees asked Washington to help build international pacing mechanisms. Elsewhere, Claude contributed to experimental cryptanalysis and a new robotics company put a claimed fleet of 100 machines behind a browser.
Korea’s AI-chip expectations outran a sixfold profit record
SK Hynix reported quarterly operating profit of 60.5 trillion won—about $41.2 billion and six times the result from a year earlier—yet the figure still fell short of forecasts. By midday on July 29, AP reported that Seoul’s KOSPI had fallen 8%, with SK Hynix down 12.6% and Samsung Electronics down 8%.
Those are dated intraday moves, not live prices. Their significance is the gap they expose between exceptional business performance and even more exceptional expectations. Memory suppliers have become critical beneficiaries of the AI buildout, but that status also makes them vulnerable when record results fail to clear an elevated consensus.
Amazon chooses concentration over Nova’s breadth
Amazon is deprecating Nova Premier, Omni, Reel and Canvas while shifting resources toward a next-generation frontier-model effort, Reuters reported, citing Business Insider. The company says it will continue supporting models customers rely on and is still investing in the next generation of frontier research.
The distinction matters. This is not Amazon abandoning AI or necessarily abandoning the Nova name; it is a portfolio reset. Maintaining separate flagship text, multimodal, image and video systems consumes scarce engineering talent and compute. Amazon appears to have concluded that a narrower attempt at the frontier offers a better chance of standing out against OpenAI, Anthropic and Google.
Frontier employees take the pacing argument international
A public letter had an observed 1,178 signatures from employees at frontier-AI organizations as of this edition. Signing in their personal capacities, they ask the United States to lead the development of international technical and governance tools that could deliberately pace advanced AI development.
The Pacing the Frontier letter is more specific than a general call to slow down. It treats pacing as a coordination problem: unilateral restraint is difficult when companies and countries fear losing ground, so credible mechanisms would need international backing. The signatures represent individuals, not formal endorsements from their employers, and the displayed count may continue to change.
Claude contributes to cryptanalysis—within important limits
Anthropic says Claude Mythos helped researchers develop a new attack on HAWK, an experimental post-quantum signature candidate, reducing its claimed key strength by roughly half. It also helped improve an existing attack on seven-round AES-128 by 200–800 times.
The headline needs its qualifiers. Anthropic’s research does not show that Claude broke the AES-128 used in production, which has 10 rounds. HAWK is not deployed, and Anthropic says no production systems are affected. The result is nevertheless notable because it moves model-assisted security research beyond finding implementation bugs and toward reasoning about the mathematical construction itself.
Enigma puts physical AI behind a browser
Enigma emerged from stealth with a $71 million seed round and says it has opened 100 real AI-powered robots to public interaction through a browser. According to the company’s release, users can direct the machines in real time, while the resulting interactions help Enigma study how people naturally communicate with robots.
The experience is designed to make physical AI participatory rather than something audiences watch in a polished launch video. It could also create a useful stream of human-interaction data, although the fleet’s scale, availability and performance are company-reported claims that will need independent observation.
The money picture is increasingly selective. Korea’s sell-off shows how much future AI demand is already embedded in hardware expectations; Amazon is concentrating compute rather than sustaining a broad model portfolio; and Enigma’s seed round shows that physical AI can still attract substantial early capital when paired with a tangible demonstration. Watch whether memory demand justifies the market’s expectations, whether Amazon’s narrower frontier effort produces a differentiated model, and whether independent researchers reproduce Anthropic’s cryptographic results. Signal, not advice.