Intel is developing a high-memory inference GPU for conventional air-cooled servers. The larger question is whether hardware like this could make existing data center capacity economically useful for more AI workloads.
An Anthropic model filed a false police tip during a web test. The challenge is to stop unauthorized submissions without stopping AI assistants from helping people make legitimate ones.
Meta's latest open-weight release enters a field where Chinese models have attracted substantial downstream development. But the companies building popular models are not necessarily the ones paid to run them.
On-device AI is making personal activity searchable across phones, PCs, and smart glasses. But shared access to a record does not ensure that every AI application will know when the information it relies on has changed.
Federal retirement could take months. A one-week test finished one real case in two days by tracing every handoff—and revealed what the software actually had to replace.
Reinforcement learning environments let models repeat multi-step software tasks under controlled conditions. For AI training-data companies, finding tasks that still expose a model's weaknesses is becoming part of producing the next round of training material.
Sleepmaxxing treats sleep as something to optimize. Smart-ring maker Oura and sleep-tech firm Eight Sleep turn part of that work into a service: the former interprets sleep data and gives guidance, while the latter uses those data to adjust the bed overnight.
We have not stopped carrying things out; we may be losing the capacity to act.
Specialized models are making machine judgment cheap enough to spread. The harder work is defining the rules they should use.
Meta’s Muse can create its own connectors—software links that allow an agent to use external services. If agents can increasingly build these bridges themselves, the true value lies less in the connector itself and more in the access behind it.
New ways of evaluating AI agents are beginning to measure when people have to step in, how often they are needed, and how much that human involvement costs.
On Chinese social media, concepts such as subjectivity, a sense of boundaries and emotional value increasingly offer explanations for ordinary experiences. Their portability makes them useful for making sense of life, but also easy to reuse as content until the explanation itself becomes a cliché.
Data Centers
As AI data centers request enormous amounts of future electricity, grid operators increasingly need to judge which planned data centers are likely to actually be built before including their expected electricity use in future grid plans.
AI Agent
LLM-based agent simulation systems can keep a modeled population available beyond one study, allowing researchers to return with another question or another scenario. That lowers the cost of testing ideas—and turns validation into a continuing task.
On-device AI
John Ternus’s first keynote suggests that Apple may see personal AI less as a new device than as something spread across devices—with the iPhone still at the center.
AI Agent
Websites are beginning to let agents act directly instead of making them work through the same steps people do.
AI Companions
AI companions are turning memory into infrastructure—and making relationships surprisingly difficult to move.
Physical AI
Robot training is moving out of the lab and into factories, warehouses, homes, and other real-world settings, making access to everyday tasks part of the physical AI race.
Tokens
Tokens are becoming industrial metrics before we agree on how well they capture useful output—and robot-hours show the same process beginning in robotics.