AI Daily Digest — Oct 24, 2026

TL;DR

4 takeaways

Block unveils Dessa, a full-duplex speech-to-speech model it says is trained entirely on human speech and runs locally on a MacBook M4 Max or better; no public access yet
Sources: TechCrunch
Anthropic and Theorem say Claude produced machine-checked proofs of no-crash properties for 239 entry points across 123 open-source libraries
Sources: Anthropic
Google Cloud announces Agent Compute Engine, the hardware-isolated microVM service it says already powers Gemini agents, with preview planned in the coming months
Sources: Google Cloud
Short: Moonshot AI releases Kimi K3.1 Pro, with full weights promised by Nov 4
Sources: Moonshot AI

Story 01 / 03

What happened: Block’s research lab unveiled Dessa on October 23, a full-duplex speech-to-speech model that can listen and talk at the same time. Block says it was trained entirely on human speech using a single NVIDIA GPU rack, rather than routing audio through text and a separate language model, according to TechCrunch. Block’s researchers posted sample conversations with laughs, pauses and shifting emotion.
Why it matters: Block says the model can run locally on consumer hardware such as a MacBook M4 Max or better. There is no public access yet; TechCrunch reports Block has asked safety researchers to test it first.
Enterprise impact: On-device voice could lower the cost of voice ordering and support flows; Square already uses voice AI for ordering, TechCrunch notes.
Startup impact: A local full-duplex model would remove per-minute cloud speech costs, if and when it becomes available.
What to focus on: Whether and how Block opens access after safety testing.
Source: TechCrunch(opens in a new tab)

Comment with your GitHub account. One thread per story.

Story 02 / 03

What happened: Anthropic published research with Theorem showing Claude proved that 239 OSS-Fuzz entry points across 123 open-source libraries always exit normally without memory-safety errors or failed assertions, for every input under 64 MB. The proofs work on compiled RISC-V binaries and total about 180 million lines of Rocq, according to Anthropic, and Claude found and patched bugs along the way.
Why it matters: Anthropic calls it, to its knowledge, the first large-scale effort to use AI for assembly-level formal verification of production software. The guarantees hold only under stated assumptions: a single-threaded static binary with a fixed memory budget on a RISC-V CPU running Linux.
Enterprise impact: Security teams get a preview of proof-backed guarantees as a complement to testing and fuzzing.
Startup impact: Theorem is offering to run the pipeline for other open-source projects and is publishing a Claude skill for developers.
What to focus on: Extending the proofs to more CPU architectures and operating-system behavior, which Anthropic says is within reach.
Source: Anthropic(opens in a new tab)

Comment with your GitHub account. One thread per story.

Story 03 / 03

What happened: Google Cloud announced Agent Compute Engine at AGNTCon + MCPCon, a compute service it says already powers Gemini as an agent for work tasks. Each agent gets its own hardware-isolated microVM that boots quickly, pauses while the agent waits, and resumes from snapshots. Preview is planned in the coming months. Google also extended Gemini for business agent governance to agents wherever they run.
Why it matters: Google says each microVM runs with its own short-lived identity and that service credentials are injected outside the sandbox, a design aimed at running untrusted agent code safely.
Enterprise impact: Teams running code-executing agents get a managed isolation option from a major cloud.
Startup impact: Agent builders could offload sandboxing instead of building their own microVM fleet.
What to focus on: Preview timing and pricing.
Source: Google Cloud(opens in a new tab)

Comment with your GitHub account. One thread per story.

Kimi K3.1 Pro
Moonshot AI’s new model has a 1M-token context and a 2 to 3x faster Highspeed variant; Moonshot’s own figures put it at 65.7% on Terminal-Bench 4.0. Full weights are promised by Nov 4.
Moonshot AI(opens in a new tab)

Proof, not just tests

Anthropic’s verification work hints at a future where AI writes the security guarantees, not only the code.