Self-Building AI, a $115B Chip Deal, and the Botnet Hidden in Your TV Box

Anthropic has just warned that AI capable of rewriting itself may be closer than anyone expected, and that single headline sits at the centre of a week where chip money, security failures, and AI ethics all collided at once. Here is what happened, and why it matters if you are learning to work in tech.

Anthropic says self-improving AI is nearly within reach

Recursive self-improvement (RSI) is the theoretical point at which an AI system becomes capable enough to meaningfully train and improve its own successors, potentially setting off rapid, hard-to-control capability jumps. Anthropic has now stated publicly that it believes it is very close to that threshold. For anyone studying machine learning or AI safety, this is not abstract: it means the assumptions baked into today's training pipelines, safety benchmarks, and deployment practices could need rethinking on a fast timeline. Read Anthropic's RSI warning on TechCrunch.

Separately, Anthropic quietly updated Claude Code's system prompt to stop users from having the model clone existing video games wholesale. The move follows a surge of AI-generated ports and decompilations flooding retro gaming communities. Importantly, Ars Technica points out that one model maker's policy change is unlikely to stem the broader wave, since many other tools remain available. The story is a clear example of why copyright and AI policy will keep colliding as code-generation tools get more powerful. See how Claude's new rules work on The Verge and why the cloning boom will continue on Ars Technica.

The $115 billion chip deal reshaping AI infrastructure

Marvell Technology has agreed to acquire optical networking company Lumentum for $115 billion, making it one of the largest acquisitions in tech history. Both companies share a notable common investor: Nvidia. The deal reflects how much pressure the AI boom is putting on the physical infrastructure underneath it, specifically the chips and photonics hardware that move data between servers at speed. If you are curious about where the money actually flows in the AI industry, this merger is a useful case study. Get the full breakdown on TechCrunch.

An Android manufacturer secretly built a botnet into 1.3 million devices

Google has revealed that one of its certified Android manufacturing partners, Shenzhen-based Videostrong Technology, installed a hidden backdoor across 1.3 million set-top boxes and used them to form a botnet. Google says it has dismantled the backdoor and permanently banned the manufacturer from the Android ecosystem. For anyone studying cybersecurity, the case is a stark illustration of supply-chain risk: a device that carries a trusted brand's operating system can still be compromised at the hardware or firmware level long before it reaches a consumer. Full story on The Verge.

Spyware warnings hit German politicians' iPhones

Apple sent threat notifications to German lawmakers and at least one filmmaker, alerting them that their iPhones may have been targeted with spyware. Germany's government is treating the warnings seriously. The incident reinforces why mobile security is a genuine professional concern, not just a talking point, and why understanding device-level threat models is increasingly relevant for security practitioners. Read the report on TechCrunch.

AI model training without the PhD: Trajectory raises $45M

Trajectory, a one-year-old startup now valued at $300 million, has closed a $45 million funding round with a specific goal: letting people who are not machine learning engineers train their own AI models. The pitch matters because one of the biggest gaps in enterprise AI adoption right now is the shortage of people who can customise foundation models for specific tasks. Tools that lower that barrier could meaningfully change what a non-specialist developer or data analyst can accomplish. Details on TechCrunch.

Google's Dreambeans expands AI storytelling to 13 Indian languages

Google's experimental Dreambeans app, which generates personalised narrative stories from user inputs, has expanded to India with support for 13 local languages. The move is notable for two reasons. First, it shows that major AI labs are actively prioritising multilingual and low-resource-language support, an area where the gap between English and everything else has historically been wide. Second, for developers interested in natural language processing, this kind of rollout signals where product investment and research energy is heading. Story on TechCrunch.

Apple's hardware squeeze: record iPhone sales, scarce supply, soaring chip costs

Apple posted record iPhone revenue in its latest quarter, up 17 percent year-on-year, despite not releasing a standard iPhone model during the period. The growth came largely from the iPhone Duo, which remains in short supply. CEO John Ternus confirmed the company cannot yet predict when production will catch up with demand. Meanwhile, Apple's inventory has surged dramatically as memory chip costs climb, raising real questions about whether future device prices will follow. For learners tracking the hardware side of tech, this illustrates how semiconductor economics ripple from the factory floor all the way to the consumer price tag. iPhone sales record on The Verge | Supply constraints on TechCrunch | Memory chip cost pressures on TechCrunch.

Apple hints at an AI revenue play, and a home robot for 2027

On its earnings call, Apple gestured at new revenue opportunities tied to third-party AI apps, without specifying a model. Given Apple's existing App Store infrastructure, the most likely avenues involve some form of platform fee or integration deal, but the company kept its cards close. On a different front, Ars Technica reports that Apple's home robot project, internally backed by CEO John Ternus, is targeting a roughly $1,000 price point and a 2027 launch, with initial use cases that are fairly modest: FaceTime calls and smart-home controls. Both stories point toward Apple positioning hardware as the trusted surface through which AI interactions happen. Apple AI revenue hints on TechCrunch | Home robot details on Ars Technica.

OpenAI wants its data centres to double as emergency power reserves

OpenAI has announced a pilot programme called Grid Protectors, starting at a planned data centre in Georgia. The idea is that during extreme weather or grid emergencies, the data centre can release up to 1,000 megawatts back to the local grid, enough to run air conditioning in hundreds of thousands of homes. It is a smart piece of public-relations positioning at a time when AI infrastructure's enormous energy appetite is drawing criticism, and it also hints at an emerging model where large compute facilities act as grid-scale batteries or demand-response assets. Full announcement coverage on The Verge.

Squarespace cuts nearly a fifth of its workforce

Squarespace has laid off 19 percent of its employees as part of a restructuring. The company, which went private in 2023, has not detailed which teams were most affected. For those entering the web development space, it is a reminder that platform businesses face real pressure as AI-assisted site builders lower the cost of entry and intensify competition. Report on TechCrunch.

Media outlets sue Brave over alleged content scraping

The New York Times and other publishers have filed a lawsuit against Brave, the privacy-focused browser, alleging that it scraped copyrighted content and distributed it to customers including Anthropic and Perplexity. The case adds to a growing stack of litigation shaping how AI companies can legally source training and retrieval data. Anyone building products that aggregate or summarise third-party content should be watching this space carefully. Full details on TechCrunch.

The next few months look set to bring more clarity on several of these fronts, from court decisions on AI and copyright to whether Apple's supply chain can keep pace with demand. Each development is another data point in understanding how the industry is actually being built, which is exactly the knowledge that separates a confident tech professional from someone just reading headlines.