Claude Code is now hackable with small TypeScript mods, Meta's Muse agent is moving onto smart glasses, and NVIDIA gave its local AI box a new memory tier with a new price. Anthropic's IPO slipped again according to Bloomberg, Tesla beat Q3 delivery estimates, and Microsoft shipped three voice models in a single day. Seven stories, sources at the bottom of each.
Anthropic opens up Claude Code with small TypeScript mods
Over October 1 and 2, Anthropic shipped a feature called mods: small TypeScript functions that change how Claude Code behaves. A mod can run before, after, or in place of events like tool calls, permission requests, prompt submissions, and UI rendering. Mods ship as plugins, install with /plugin on both the CLI and the desktop app, and require Claude Code v2.1.287 or newer.
— Anthropic · TechDig
The surface area is wide. Mods can rewrite prompts, block or retry tool calls, approve or deny permission requests, strip secrets from tool output, and draw custom UI: buttons, panels, /commands. Anthropic is also migrating its own built-ins into mods. /diff is already a mod you can disable or replace, and AGENTS.md loading plus telemetry are built-in mods too, with more to follow. The three official examples give a feel for the range: Token Weather puts a live context window forecast above your prompt, Blast Radius dry-runs dangerous commands like rm -rf and previews the blast area first, and Replay Theater steps you through a round of edits one change at a time.
Here is the part worth reading twice: mods are not sandboxed. They run with the same permissions as Claude Code itself, meaning they can read secrets, rewrite prompts, and approve tool calls before you see them. Enterprise installs get a built-in sec-default mod that loads first and blocks any attempt to override permission-deny rules. Anthropic uses Claude to write mods as well, so the agent can generate the TypeScript, install it, and hot reload it. Extensibility this deep is genuinely useful, but the trust boundary is now whoever wrote the last mod you installed.
Meta is putting its Muse agent on smart glasses
On October 2, Meta announced that its AI agent Muse is coming to smart glasses soon. You wake it by voice, with Hey Muse or a custom name, leave the phone in your pocket, and tasks run in the background until they finish and report back.
— Meta · TechCrunch
Muse runs on the MuseSpark model, and its core execution environment is a private, persistent Linux VM in the cloud with its own browser, filesystem, and terminal. Close the app and the work keeps going. The agent writes and executes its own code, dispatches multiple subagents, and sets up scheduled tasks, producing documents, PDFs, websites, and dashboards along the way. For glasses specifically, Meta lists reservations, flight searches, finding parts, logging meals, workout guidance, and buying the item right in front of you. Payments go through Stripe Link one-time virtual cards, so the AI never touches your actual bank card details.
The privacy terms deserve a close read. Muse data is not used for advertising, but it is used to train future models unless you actively opt out. A Confidential VM that even Meta cannot access is promised later this year. Context matters here: Muse launched on September 8, and Connect 2026 positioned it as the center of Meta's consumer AI push. Hardware you wear all day plus a VM that never sleeps is the most concrete consumer agent deployment yet, and the default-on training data is the price of admission.
NVIDIA adds a 64GB DGX Spark tier at $4,999
On October 2, NVIDIA announced a new 64GB unified memory configuration for DGX Spark, starting at $4,999 with a ship date of October 23. It is sold through OEMs including Acer, ASUS, Dell, Gigabyte, HP, and MSI. There is no Founders Edition for this tier.
— NVIDIA · Tom's Hardware
The hardware story is unchanged where it counts: the GB10 Grace Blackwell superchip, DGX OS, the full CUDA software stack, and ConnectX-7 all carry over. Of the 64GB, about 8GB is reserved for the system, leaving roughly 56GB for weights plus KV cache. NVIDIA's official line is that models up to 100 billion parameters run locally. Under NVFP4, Qwen3.8-27B weights take 16.5GB, leaving 40GB for KV cache, and Gemma4 26B-A4B fits in 15GB. Two 64GB units connected directly over QSFP form a 128GB memory pool with a 200 billion parameter ceiling, and Qwen3.8-27B measured up to 1.7x throughput on that two-box cluster. The Sync Cluster Assistant can auto-network up to four units.
The pricing moves are the real story. The 128GB Founders Edition goes up to $6,950 effective October 2, after launching at $3,999 in 2025 and rising to $4,699 in February 2026; media reports attribute the increase to constrained memory supply and rising costs. NVIDIA also claims up to 1.9x faster local agent inference from deep optimizations to llama.cpp and vLLM. Read together: the entry price came down, the top price went up, and the sales pitch is now explicitly local agent workflows.
🔗 ITHome
Tesla Q3 deliveries beat expectations, storage misses
Tesla's IR press release on October 2 put Q3 production at 464,391 vehicles and deliveries at 486,532, split into 478,237 Model 3/Y and 8,295 other models. Energy storage deployments came in at 13.7 GWh.
— Tesla IR · CNBC
Deliveries beat the market expectation of roughly 463,761 units, about 5% above, though the company-compiled consensus sat at 461,974. The quarter still fell short of the Q3 2025 record of 497,099, down 2.1% year over year, but grew sequentially from 480,126. The stock rose more than 5% at one point intraday. Storage is the weaker column: 13.7 GWh missed the roughly 15.9 GWh consensus, even though it edged past Q2's 13.5 GWh and last year's 12.5 GWh.
The same week's SEC filings add financing detail: $30 billion in new loans and credit facilities to support the expansion, with capital expenditure above $25 billion this year across factory buildout and Robotaxi. Cybercab, the two-seater with no steering wheel and no pedals, has started production, and the full earnings report lands after the close on October 21. A delivery beat is nice, but the debt and capex numbers tell you the Robotaxi bet is being financed in real time.
🔗 Tesla IR
Microsoft ships three voice models in one day
On October 1, the Microsoft AI blog announced three models at once: MAI-Transcribe-2-Streaming for streaming speech to text, plus MAI-Voice-2.1 and MAI-Voice-2.1-Flash.
— Microsoft AI · Artificial Analysis
MAI-Transcribe-2-Streaming takes first place on Artificial Analysis' 38-model leaderboard for final transcription WER at 2.5%, ahead of the previous leader Grok Voice Transcribe 2.0 at 2.7%, with latency of 0.13s against 0.49s. First partial results arrive within about 100 milliseconds of receiving audio; the first partial transcript posts about 0.12s at 2.5% WER. The model auto-detects 60 languages. Pricing is $0.54 per hour under a promotion through year end, about $9 per 1,000 minutes, which is above ElevenLabs Scribe v2 Realtime and Deepgram Flux at roughly $6.50. Microsoft is competing on quality, latency, and enterprise-grade deployment rather than on price.
MAI-Voice-2.1 covers 23 languages across 26 regions and keeps a single voice consistent across languages while using local accents. It costs $22 per million characters. The Flash version is $15 per million characters, runs inference 55% faster, and comes in about 60% cheaper than comparable models. Both support voice cloning from a few seconds of reference audio with a consent mechanism. Microsoft's in-house model family, including MAI-Transcribe-2, MAI-Thinking-1, MAI-Code-1.1-Flash, and the MAI-Image series, keeps growing and plugs into Foundry, Copilot, and Teams. Second on price, first on WER is a deliberate position, not an accident.
Spanner Queues reaches general availability
Google announced that Spanner Queues is now generally available. It provides transactional messaging inside a Spanner database, aimed at asynchronous tasks in agent workflows.
— Google Cloud
Messages commit and acknowledge atomically with database writes in the same transaction. There is a pull model with a SQL interface, and messages are stored as rows, so you can query, join, and filter them directly. You can schedule sends with a future timestamp, extend leases, and dead letters keep retrying until they are acknowledged or deleted. The pitch is straightforward: no separate queue infrastructure to provision, no Kafka-class system to operate, scaling with the Spanner instance and inheriting its cross-region availability.
The direction is worth naming: agent workflow infrastructure keeps getting absorbed into core databases, and the money flow says the same thing. For teams already paying for Spanner, one fewer system to run is a real argument. The open question is how the throughput profile compares to a dedicated broker once agent traffic gets heavy, and that answer will only come from production.
Anthropic's IPO roadshow reportedly slips to November
According to Bloomberg, Anthropic's IPO roadshow is now targeted for the week of November 9, with listing before Thanksgiving on November 26. That is a second delay within two months, after September reports pointed to mid-October ahead of the midterm elections. The company declined to comment.
— Anthropic S-1 · Bloomberg
Potential investors are discussing a valuation between $1.8 and $2 trillion, nearly double the $965 billion private round in May. From the filed S-1, the hard numbers are 2025 revenue of $4.59 billion and compute contract obligations of $518 billion, most of it rigid spend owed whether or not the capacity is used.
Keep the sourcing straight: the roadshow timeline is attributed to Bloomberg's sources, not an official announcement, so treat it as reported intent. The number to sit with is the gap between a $4.59 billion revenue base and $518 billion in compute commitments; whatever valuation the market accepts has to price that difference. Whether this date holds is the next data point.
KD Agentic · AI Daily Digest
Top comments (0)