AI Daily Digest October 18, 2026: Reflection Opens Beam 501B, Anthropic Replaces Git, TSMC Rides the AI Wave

Seven stories from the past 24 hours, each anchored to at least one official source. This edition keeps circling one theme: when agents do the work, everything around the model, from version control to sandboxes to the checkout counter, has to be rebuilt around them.

1. Reflection Ships Beam's Weights: 501B Parameters Under Apache 2.0

Reflection released the full weights of Beam on Friday, eleven days after previewing it, and the license is the permissive one: Apache 2.0, on Hugging Face, with no restrictions on distillation, fine-tuning, or commercial deployment. Beam is a sparse mixture-of-experts model with 501 billion total parameters that activates 23 billion per token, holds a 1M-token context window, and was pretrained on 23.8 trillion tokens. The reinforcement learning stage ran on 10,500 NVIDIA GB300 GPUs for four weeks and produced more than 100 million rollouts.
— Reflection · Hugging Face

Reflection's own benchmark sheet puts BF16 Beam at 81.6% on SWE-bench Verified, 80.1% on Terminal-Bench 2.1, 90.6% on GPQA Diamond, and 97.7% on AIME 2026, and claims it matches leading open models while using roughly a third of the reasoning tokens. Independent scoring landed the same day at 30 on the Artificial Analysis Intelligence Index, behind GLM-5.3 at 45, Kimi K3 at 44, and Mistral Large 4 at 38, so the honest read is a strong coding model whose reasoning rank is still unproven. Weights ship in BF16, FP8, and NVFP4 precision plus a DFlash2 speculative decoding drafter; the NVFP4 build compresses roughly 1TB of parameters down to about 280GB, which fits four B200s with room for context.
— Reflection · Artificial Analysis

The part that outlives the benchmark argument is the packaging. Reflection launched with more than 40 partners forming what it calls an open operating system for agents: day-zero support in Transformers, vLLM, SGLang, Unsloth, and NVIDIA Dynamo, hosted endpoints on AWS, Azure, Google Cloud, Baseten, CoreWeave, Together, and OpenRouter, and on-premises delivery through Dell and HPE. Its own Mirror CLI harness, a Rust core with a Python frontend, was open-sourced alongside the weights, and the first-party API is in a free public beta. Reflection also wrote a $10,000 Hugging Face credit bounty for the best community projects. With the Financial Times reporting unconfirmed acquisition talks with NVIDIA, shipping the weights now looks like a deliberate move to make Beam everywhere before anything else is decided.
— Reflection · Hugging Face

🔗 Reflection Blog · Hugging Face

2. Anthropic Replaces Git: 30,000 Merges a Day Broke It

Anthropic published an engineering post on its new version control system, Scion, and the numbers explain why the company left git entirely. Daily merged changes grew from about 2,500 in May 2026 to more than 30,000 by September, a tenfold jump in four months driven mostly by Claude writing code. The monorepo carries almost 100GB of main-branch history, a compressed working copy exceeds 25GB, and CI runs about 25 million checkouts per day, which would mean over 600 petabytes of daily git traffic if fetched the naive way.
— Anthropic Engineering

A seven-person team built Scion and completed the full migration in five months, shipping a new code review tool called Antique, code search, and source control IAM in the same window. The design inverts git's assumptions for a world where nearly every client is an agent that is always online: commits, merges, and rebases happen server-side as changesets in a write-ahead log, there are no packfiles to maintain, checkouts are sparse and virtualized by default, and any job can query file history through an API instead of holding a local copy. The result, by Anthropic's figures, is infrastructure that uses 200 times less CPU and memory than the old git mirror fleet, with checkouts dropping from minutes to seconds and CI reclaiming 25% of its CPU capacity.
— Anthropic Engineering

The honest detail is the human one. Anthropic says Claude wrote much of Scion itself, and in one example an internal snapshot of Claude Mythos Preview rewrote a prototype ingestion pipeline to use 20 times less memory while running four times faster, but engineers describe their main job as telling Claude no when it violated core constraints or over-engineered a solution. The company is explicit that Scion is not a general git replacement and will not be released publicly; the design lessons go to GitHub, GitLab, and the open source community instead. Its prediction is that most organizations running agentic coding will hit the same wall within six to twelve months.
— Anthropic Engineering

🔗 Anthropic Engineering

3. Perplexity Open-Sources SPACE, the Sandbox That Kept Nine Frontier Models Contained

Perplexity open-sourced SPACE, the sandbox platform behind its Computer agent, under Apache 2.0. SPACE, short for Sandboxed Platform for Agentic Code Execution, runs agent code inside Firecracker microVMs behind an HTTP API. Since July it has replaced a commercial sandbox provider and powered every Computer session, handling millions of sandbox creations per day, with creation running three to five times faster than the previous setup on the same production traffic.
— Perplexity

The security claims deserve both attention and caution. Perplexity says that across 118 escape attempts by nine frontier models, none broke out of SPACE, while the same models could escape from other sandboxes that are not open-sourced. An earlier published evaluation found no VM-to-host escape in 108 attempts but did surface network restriction bypasses, which the team fixed before release; it also disclosed related weaknesses in competing sandbox products to their vendors. Every one of these numbers is Perplexity's own claim, not an independent audit, which is exactly why open-sourcing the code matters: now anyone can run the tests.
— Perplexity · Semafor

The architecture decisions are the transferable part. SPACE treats each sandbox as untrusted and manages network access and platform-held credentials outside the agent's working environment, so agents can call authorized services without ever holding the keys. Frequent snapshots and full-environment checkpoints preserve progress across interruptions and make session branching cheap, supporting flows like pausing for human review and resuming with state intact. Distribution goes through CoreWeave, Akamai, and Crusoe, which matters because most enterprises will rent a hosted sandbox rather than operate Firecracker fleets themselves.
— Perplexity

🔗 Perplexity Blog · GitHub

4. Google Research Puts AI Agents on Wet-Lab Robots

Google Research published a post on grounding AI agents in physical wet labs, and the centerpiece is VERID, a semantic digital twin of the laboratory workcell built on the open source PyLabRobot framework. VERID maintains a live, stateful 3D volumetric entity tree covering robotic arm geometry, well volumes, reagent identities, and pipette tip inventory. Agents plan, simulate, and execute liquid-handling protocols with four safeguards: live 3D state tracking, fast virtual dry runs that catch collisions in seconds, hardware safety checks that block impossible commands, and synchronized video plus action logs for reproducibility.
— Google Research

The closed-loop results are what make this more than an infrastructure demo. In a self-play campaign, an agent reconstructed blinded dye mixtures from overhead camera images and identified its own optical and fluidic artifacts, improving accuracy more than fivefold. In aptamer discovery, where a full design-test-analyze cycle takes seven to ten days, agents pushed confirmed on-target hit rates to 33.5% for myoglobin and 42.1% for alpha-thrombin over three rounds. The first round on alpha-thrombin produced zero binders because fixed primer overhangs disrupted folding in 48% of sequences, and the agents revised their own design rules after seeing the data, which is the entire point of lab-in-the-loop learning.
— Google Research

The protein binder work points at drug discovery directly. Applying the same workflow to FAP, a receptor overexpressed in solid tumors, yielded binders with single-digit nanomolar affinity in just a few experimental runs, and an epitope excision technique that folds only the local binding pocket accelerated each cycle by a factor of 8.6. Human-agent collaboration, where a scientist sets strategy and the agent drives execution, nearly doubled the in-silico hit rate compared with a scientist working alone. Google is careful to frame all of this as prototype research on deliberately non-infectious systems, aligned with its Frontier Safety Framework, not a shipped product.
— Google Research

🔗 Google Research

5. Replit Makes ChatGPT Its Default, and the Anthropic Era Quietly Ends

Replit CEO Amjad Masad said on stage at TechCrunch Disrupt that Replit is no longer working with Anthropic, citing cost, partnership terms, and performance. His benchmark claim was specific: on Replit's public benchmark for building vibe-coded applications from scratch, OpenAI's Sol and Astra consistently outperform. The company announced an OpenAI partnership in August, but this was the first public signal that Claude had lost its seat at the center of Replit Agent.
— Replit CEO · TechCrunch

The follow-up clarified the real shape of the change. Replit remains model agnostic and customers can still select Claude, but OpenAI is now the default powering Replit Agent, and defaults are where the money is: most users never change them. Replit had been a top-five Anthropic customer until recently, so the spending shift is substantial. Masad also laid out the business mix behind the move, with enterprise contracts growing from 2% of revenue last year to 15% this year, more than 50% of revenue from B2B, and a prediction that enterprise and government will exceed half of revenue in 2027.
— Replit CEO · TechCrunch

Not everyone accepts the clean-break narrative. A person familiar with the partnership told TechCrunch it was still in place, and a September Replit blog post cited Claude Fable 5 and 5.1 running in Replit Agent's production environment. Masad's own framing was provisional, ending his on-stage answer with for now. The more durable lesson sits in a separate on-stage remark about deployment: agents need a habitat, meaning orchestration, error recovery, sandboxing, networking, and identity. The platform economics and the infrastructure checklist are converging on the same conclusion, that the model is one component among many.
— Replit CEO · TechCrunch

🔗 TechCrunch

6. TSMC's Q3: The AI Buildout Shows Up in the P&L

TSMC reported third-quarter results on Friday and the AI infrastructure boom is no longer a forecast on this income statement, it is the income statement. Net profit rose 69% year over year to $23.6 billion on revenue up 51% to $47 billion, and the company raised its 2026 capital expenditure guidance to as much as $64 billion, a record. Advanced nodes running AI accelerators remain the growth engine, and the foundry is spending like the demand is durable.
— TSMC · Reuters

The capex number is the one to watch, because it propagates. TSMC's spending becomes orders for equipment makers, construction firms, and power providers, and its yield on each new node sets the effective supply curve for every GPU and custom accelerator shipped next year. When a single supplier posts 51% revenue growth and still cannot build capacity fast enough, that tells you the constraint in the AI economy has moved from model design to silicon and the fabs that etch it.
— TSMC · Reuters

For the broader market the results land amid rising scrutiny of AI circular financing, where chipmakers and neoclouds fund the very demand that buys their products. TSMC sits upstream of most of that loop, selling manufacturing capacity rather than financing purchases, which makes its margin and utilization figures one of the cleaner signals available for separating real end demand from vendor-financed demand. This quarter, that signal says the demand is real.
— TSMC · Reuters

🔗 TSMC · Reuters

7. Google's Universal Cart Turns Search Into a Checkout Counter

Google made Universal Cart available to US users, closing the loop between product discovery and purchase. The tool collects items found in Search and the Gemini app into a single cart that works across retailers, sends alerts when a saved item drops in price or comes back in stock, and factors credit card perks and promo codes into which payment method it recommends at checkout.
— Google · TechCrunch

Checkout runs on Google's Universal Commerce Protocol, an open standard introduced earlier this year to let AI agents and retailers transact without custom integrations on both ends. First previewed at Google I/O in May, Universal Cart is now the first consumer-scale deployment of that protocol, which means agentic commerce has moved from conference demos to daily traffic. The design is deliberately unglamorous: a cart, some alerts, and a payment optimizer, rather than a shopping agent that bargains on your behalf.
— Google · TechCrunch

The strategic read is about defaults again. Whoever holds the cart holds the merchant relationship data, the payment routing, and the moment of purchase intent, and Google is wiring that position directly into both Search and Gemini. For retailers the protocol promise is standardization, but the practical question is whether an open spec with Google as its largest operator behaves like open infrastructure or like a storefront with extra steps.
— Google · TechCrunch

🔗 Google · TechCrunch

KD Agentic · AI Daily Digest