Daily AI Digest

Activity Frames compiles screen activity into deterministic, replayable memory without a model

Activity Frames uses a zero-model pipeline to segment local screen capture into records with application, site, timing, input volume and pointers to raw evidence. The output is cacheable, auditable and byte-identical for the same input. On 128,756 frames across 51 active days from one professional, it compressed a day into a prompt block 86 times smaller in 68 milliseconds.

An agent reading the block answered daily-activity questions at 98.4% accuracy, versus 66% to 80% after LLM summarization. The paper also demonstrates zero-model-token replay on a matched compiled routine. Every result comes from one user, so privacy, cross-person generalization and real-task value remain open.

Verdict: Teams building local personal agents or auditable memory should care. Compiling repeated human activity into deterministic state may be more controllable than endlessly expanding model-written summaries.

DataSpace tests agents across 7,439 mixed artifacts, where the best accuracy is still 66.34%

DataSpace contains 410 cross-language tasks and 7,439 artifacts totaling 15.01 GB across CSV, JSON, SQLite, Markdown, PDF and video. An agent receives only a question and local workspace, then must return the complete table. A deterministic evaluator normalizes columns, types, precision and row order.

Across six frontier multimodal models and five common agent harnesses, the best accuracy reached 66.34%. Holding the backbone fixed, harness choice produced a 15.36-point spread, while multimodal evidence and joins reduced performance consistently. The numbers are author-run but show that data-agent reliability is not determined by model name alone.

Box CEO Aaron Levie says working with an agent resembles managing someone inside a process rather than asking a chatbot questions. A prompt should look like a specification with scope and a definition of done. The larger gain comes from rebuilding data access, cross-team handoffs and human review instead of inserting agents into an old Q&A surface.

Product advisor Peter Yang argues that consumer AI depends less on having the strongest model than on whether ordinary users trust it with email, calendars and other apps, and understand what it can now do. For ChatGPT and Gemini, onboarding, permission explanations and reliable completion may matter more than leaderboard rank.

Vercel CEO Guillermo Rauch argues that developer tools should be both open source and universally extensible because AI coding agents are becoming the industry's most important tool category. A common plugin standard could let one extension reach CLIs, IDEs, cloud agents and personal assistants without separate integrations.

Basis cofounder Mitch Troyanovsky says perfect outcomes still do not prove a reliable agent process

Basis runs agents for hours or days on accounting and tax work. Troyanovsky says long-horizon work is hard because models lack real short- and long-term memory and cannot rely on cheap runtime feedback like code. His team translates expert process into a small set of behavior specs, such as returning to primary tax law or rendering a presentation before delivery.

The specs do not prescribe every step. A reviewing agent uses them to judge whether a trajectory followed the critical process. Troyanovsky treats context as runtime training data and says instructions and tool descriptions directly affect performance, sometimes making the English more precious than code structure. He also expects stronger models eventually to absorb much of this harness work.