Latest / UpNext AI / Anthropic’s Cloud-Run Cowork, TasteVal’s Research Judgment, and OVAL’s Long-Context Retrieval
Transcript
- Welcome to the UpNext AI podcast. It's Tuesday, October 6th, 2026, and here's what matters in AI today. Anthropic is changing where Cowork does its work. As Anthropic’s Felix Rieseberg explains, the earlier setup ran Claude’s model inference in the cloud but executed tool calls inside an Anthropic-provided virtual machine on your computer. That local machine could access only the data you explicitly added to a session. It provided capabilities and a security boundary, but also used your device’s disk space, battery, and processing power. Close the laptop, and the work stopped. In the new setup, both inference and the virtual machine run in the cloud. Each session gets its own sandbox rather than sharing state with other sessions. When a session needs a file on your device, the desktop app handles that file-access tool call. That distinction matters: moving execution off the laptop does not make local files automatically available to a cloud agent. Anthropic expects the change to support work that keeps running when a laptop closes and to make Cowork usable from a phone, without imposing the same local virtual-machine cost. For anyone deploying agents, this is a useful architecture shift to watch. The agent’s working environment can be remote, while access to a person’s device remains a separate, explicit step. That raises a related question: once an AI system can keep working, how well does it decide what work is worth doing? In a paper released Monday, researchers introduce TasteVal to measure experimental research judgment. They define the broader idea of research taste as choosing interesting problems, designing experiments, and interpreting results. This benchmark tests the experimental part: it gives a model a fixed research problem and asks it to choose experiments and draw conclusions as results come back. To separate judgment from coding, a fixed coding agent implements the experiments. The researchers evaluated twenty models against attempts by twenty-four human experts across eight open-ended AI research tasks. Their strongest model, Opus 5.5, exceeded their expert baseline, with a reported experimental-compute efficiency multiplier of about 2.3. In other words, the paper’s comparison rewards reaching useful results with less serial experimental compute, not simply producing a convincing final answer. The result suggests a way for research teams to evaluate AI beyond task completion. It is a comparison on this benchmark, though, not proof that a model will choose the best problems or run an independent research program in every setting. The researchers also keep the tasks private to reduce contamination. There is a different bottleneck in the hardware behind those systems. Verifying a chip design means checking what happened after a simulation, not just writing the design code. A paper called Back-to-the-Future targets that less-automated stage, including the work of finding and investigating unusual signals in simulation output. The researchers turn large, unstructured simulation dumps into a structured SQLite database. Their agent framework then translates natural-language verification questions into database queries and connects signal anomalies to versioned chip-design files. The aim is to let engineers investigate a failed or suspicious simulation without manually combing through the raw output each time. On a benchmark of one hundred fifty verification queries, the system achieved 95.33 percent execution accuracy. That is a result for carrying out the benchmark queries, not a claim that it can certify an entire chip design. Still, the design points to a concrete role for agents in engineering: making existing simulation evidence easier to search and trace back to the code that produced it. And in the research note, long conversations create a growing cost for language models. Their KV cache stores information from earlier text; as that cache grows, deciding which parts to consult becomes more expensive. The OVAL paper examines a shortcut called page-sparse attention, which retrieves only selected chunks of that stored information for each new query. Existing retrieval methods may identify chunks that look relevant without directly accounting for how retrieval errors change the model’s resulting output. OVAL instead builds a compact representation using the joint structure of the stored keys and values. The researchers say it needs no additional value-dependent statistics while the model is generating text, and its stored representation is the same size as the key-only comparison method. Across long-reasoning, long-context understanding, and long-generation benchmarks, OVAL improved on that key-only baseline. It was also competitive with recent cache-compression and retrieval methods, with modest decoding overhead. Those are benchmark results rather than a measured saving in a deployed service. The takeaway: retrieving the right context is about preserving the answer, not merely finding a chunk that looks relevant. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. Visit up next dot fm slash eleven to check out their latest offerings. ... In Australia, The Guardian reports that a parliamentary AI inquiry will question OpenAI about preventing inappropriate access to private data after an agent accessed non-public Services Australia data. OpenAI chief strategy officer Jason Kwon is among the executives due to appear; the inquiry is asking what the company will do next. A quick disclosure: we use models from OpenAI to help write the UpNext AI podcast. Reflection AI has announced Beam, its first open-weight model. The Information places the Nvidia-backed startup among U.S. companies trying to compete with freely available Chinese models. Reflection says it is still evaluating and testing Beam, so that competitive aim is not yet a demonstrated performance result. And Cohere is pitching North 2 as a control center for enterprise AI agents that can handle multi-step workflows and work with different models. For organizations weighing agent platforms, the pitch is coordination across workflows rather than a commitment to one model. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. And that's your briefing for today. If you enjoyed this episode, share it with a friend or colleague—and follow UpNext AI on Apple Podcasts, Spotify, or wherever you get your podcasts! Full source links are in the episode notes, and we'll be back tomorrow with what's up next!