# This Week in AI: Robots Take a Punch, Assistants Get Long-Term Memory, and Models Keep Getting Cheaper > Source: > Published: 2026-08-18 14:14:10+00:00 *A dense week across robotics, model releases and AI research, with several labs quietly shipping capabilities that would have been headline news on their own in a slower month.* Google rolled out a phone feature that recognizes sign language in real time, letting Deaf users communicate through the camera without typing or relying on a third-party interpreter app. It is a narrow use case on paper, but it points at a broader shift: accessibility features that used to require dedicated hardware are increasingly just software running on a phone's existing camera and chip. Robotics had its loudest week in months. A startup began offering home-cleaning robots for rent at 30 dollars an hour, pricing the service directly against a human cleaner rather than against a one-time appliance purchase. Separately, a six-foot humanoid robot took part in what organizers billed as the first full-contact robot fight held in the US, a made-for-spectacle event that nonetheless signals humanoid platforms are now durable and coordinated enough to take a hit and keep moving. In Xiaomi's factory, a robot has been assigned to document its own production line for the company's blog, effectively reporting on its own manufacturing process. And a new research effort claims robots can now learn manipulation skills from roughly 170 years of aggregated human video footage, using scale in place of hand-coded demonstrations. On the assistant side, ChatGPT gained the ability to retain context from what users do outside the ChatGPT app itself, extending its memory beyond a single conversation or session. OpenAI also shipped a faster version of its GPT-5.6 model, branded Sol, claiming roughly a 14x speed improvement over the standard release. Meta, meanwhile, pushed a more capable AI agent directly onto the laptop, moving further from a chat window into a background process that can act on a user's behalf. Model releases kept pace. Tencent showed a system that turns a single text prompt into an explorable 3D environment rather than a static image or video clip. China's MiniMax open-sourced a music generation model positioned as a direct competitor to Suno. Google updated Gemini Flash with a meaningful jump in coding performance while cutting its price roughly in half, continuing the trend of frontier-adjacent capability arriving at commodity pricing. NVIDIA released a 30-billion-parameter model that activates only 3 billion parameters per query, a mixture-of-experts design aimed at cutting inference cost without cutting the model's total capacity. And China's Z.ai unveiled GLM-5.3, which reportedly performed well enough on cybersecurity-related tasks that the team delayed its open release to review the implications. [Copper Hits Fresh Record Highs as AI Data Centers Push BHP Past Iron Ore](https://startupfortune.com/copper-hits-fresh-record-highs-as-ai-data-centers-push-bhp-past-iron-ore/) BHP's copper business just out-earned its iron ore division for the first time ever, and copper hit fresh record highs on the LME and COMEX this month as AI data center buildouts strain global supply. Aluminum is climbing too, though nickel has yet to break out the same way. - [AI data centers driving copper demand surge](https://startupfortune.com/copper-hits-fresh-record-highs-as-ai-data-centers-push-bhp-past-iron-ore/) - [copper shortage pushes metal to record prices](https://startupfortune.com/copper-hits-fresh-record-highs-as-ai-data-centers-push-bhp-past-iron-ore/) Anthropic had two separate items this week. CEO Dario Amodei said he believes AI could help eliminate most human disease within a decade, one of his more specific timelines on the topic to date. Separately, Anthropic confirmed Claude has made measurable progress on the Riemann Hypothesis, one of mathematics' longest-standing unsolved problems, and that Claude will now embed an invisible signature into the content it generates, a provenance marker aimed at making AI-written text traceable after the fact. Taken together, the week reflects where the AI industry's attention has moved: less about whether a chatbot can hold a conversation, and more about whether these systems can act in the physical world, remember context over time, and get cheaper and faster to run at scale.