Oct 6 · AI Daily

Google's Call for Me feature launched on the Pixel 11 series is expected to expand to personal calling scenarios, allowing Gemini to call family and friends on behalf of users, with interface examples showing "Mom" and "John." In this version, Gemini's tasks remain restricted, mainly relaying messages or making simple inquiries, and will provide an option to decline incoming AI calls. The application update also adds write and deletion permission sections, which are still in development with no official release date yet.

Claude Code v2.1.290 has been released, adding the serverToolUses field to the turn.step hook results for mods, which lists tool calls executed by the API itself along with their IDs, names, inputs, and start and end times.

Google's Gemini "Call for Me" AI feature may expand from business calls to calls with family and friends. Android Authority discovered a "Gemini Calling" onboarding page in an APK teardown, with examples including "Call Mom and say I'm running 15 minutes late" and "Call John to ask if he's coming for dinner," with the feature likely limited to short calls similar to text messages.

Claude Projects cloud chats can now connect to designated folders approved by users on their computers. The chats still run in the cloud, only reading and writing files within the folder when tasks require it, and the rollout has begun. Author Thariq called this cloud-running, local-access mode "local hands" and stated that the capability will also come to cowork.

Airbnb CEO Brian Chesky: "We will spend way more on AI tokens this year than we expected, but that's fine because the ROI is right there.

I think it's safe to say now that AI is the best thing that's ever happened to Airbnb. I feel like we are becoming an AI-first company, and that's probably the number one explanation for our performance."

• Airbnb stock surged 15% following better-than-expected earnings

From the "CNBC Television" YouTube channel, (full video link in comments)

If you live in the terminal, this one is for you.

Take a look at how to use Codex CLI to launch tasks with voice, manage agents across projects, and explore another direction in an isolated worktree.

A tutorial showing how to build an end-to-end streaming robot learning pipeline around NVIDIA Cosmos3-DROID without downloading a 707 GB dataset locally.

SpaceX shares jumped 7.6% to their highest level since mid-June, pushing Elon Musk's net worth back above $1 trillion, an increase of $30.6 billion since last Friday, cementing his position as the world's richest person. Musk previously became the world's first trillionaire after SpaceX went public in June, losing the status days later due to declines in SpaceX and Tesla shares.

NVIDIA closed at $238.90 on October 5, up 2.12% for the day, setting a record closing high for the third consecutive trading session with a market cap of approximately $5.76 trillion, firmly maintaining its position as the world's most valuable publicly traded company.

Meta and Microsoft are cutting employee usage of Claude. Microsoft previously projected at least $1 billion in internal Claude spending this year before lowering that figure by over 33%.

The Norwegian government stated it wants to temporarily ban smart glasses in certain public spaces such as parks, schools, and gyms, citing privacy concerns especially regarding unconsented filming, and plans to submit a bill as soon as possible while having an expert panel draft permanent regulations. The government clarified it is not seeking a blanket ban; French e-Enfance received about 20 reports last month of women being filmed without consent, mostly involving Meta glasses, and Paris prosecutors have launched an investigation.

A paper by Johns Hopkins and CMU proposes harness learning, training a small model to rewrite an agent's harness code (the part that determines what the model sees and what tools it calls) based on failure reports, using the score of the new harness as the reward.

tvOS 27.2 Beta 3 code shows that Apple TV 4K (4th generation) will support Siri AI and Apple Intelligence, with system settings adding an "Improve Siri and Apple Intelligence" toggle to authorize Apple to collect text and audio data from Siri, dictation, and translation to improve foundation models.

SemiAnalysis noted that the value of subscription plans cannot be viewed in isolation: monthly fees buy a certain amount of "credits," different combinations of models and token types consume varying credits, and the credit cost ratio differs drastically from the API price ratio. Therefore, the API equivalent value of the same $200/month Claude plan varies depending on the models and workloads used.

Ethan Mollick said that over the past few weeks he has moved many complex Cowork tasks to the new version of Claude Projects. Unlike older Projects, this version supports continuous conversation and uses dedicated cloud VMs. He believes the experience is better in most aspects, though documentation is unclear. A cited Cowork team member explained: the new version moves model inference and VMs to the cloud, with each session having an independent sandbox that is destroyed when the session ends, while the desktop app only handles access to files explicitly added by users, whereas the older local VM had disk, battery, and performance overheads.

Shanghai AI Laboratory released the SkillGym paper, turning human-written agent skill files into sandbox tasks with pass-or-fail checkers, training models on execution trajectories verified to pass, and reducing reliance on in-context retrieval and instruction following.

Reflection AI released its first open-weights model, Beam, a sparse MoE architecture with 501B total parameters and 23B active per token, aimed at coding, reasoning, and agent workloads, with an effective context length of 1 million tokens.

Independent researcher Syed Anas Mohiuddin's proof-of-concept attack exploits a trust gap in MCP, allowing an agent within an attacked network to propagate malicious instructions to other internal agents. Over the past five months, organizations including Google, JPMorgan Chase, Weaviate, Rapid7, the French government's interministerial digital division, and the U.S. government confirmed related vulnerabilities, which constitute a form of prompt injection targeting specific agents rather than LLMs.

A new Epoch AI report states that OpenAI researchers' coding agent usage calculated at API list prices roughly doubles every month. Starting January 2026, the median researcher's daily spending rose from less than $1 to $601 by mid-August, or about $2 million per year, approaching or exceeding the cost of hiring a researcher. The report also notes that this growth rate is difficult to sustain. The authors further quoted Boris Cherny as saying he uses Fable for everything, and Sam Altman using ultra-fast mode throughout for prompting.

NVIDIA's Olympus is a 10-wide out-of-order server core with a base frequency of just 3.3 GHz, relying on performance per cycle and implementing SMT, with single-thread performance approaching desktop levels. Its branch prediction in SPEC CPU2026 is slightly behind AMD Zen 5 and slightly ahead of Intel Lion Cove, while floating-point tests lead both.