MelbourneSunday, 8 November 2026Morning Edition
Vol. I · No. 312
Free Press

The Daily Signal

Morning Briefing
Edition
Sun 8 Nov 2026
Intelligence on the AI Frontier
Insurers become the regulators of AI autonomy, memory chips take a bite out of the richest phone maker, and power becomes a matter of zoning.
Law & Liability · The Lead

Autonomy Now Comes With a Policy Number

An AI-native law firm waited six months for a $5 million liability policy. The day it took effect, the firm stopped supervising a fifth of its agents' work.

$5MAI liability cover
20%Work now unsupervised
16 of 20Top insurers excluding AI
$80BAnthropic run rate
82.7%Q3 VC dollars to AI

The firm has about sixty lawyers and a plan to run legal work entirely through agents by the end of next year. Its cover comes from a two-year-old specialist underwriter with $1 billion of capacity and Lloyd's behind it. Once the policy was in place, lawyer review was removed from "gray-area" tasks, about 20% of the firm's work.

The rest of the market is moving the other way. At least 16 of the 20 largest liability insurers have filed to exclude generative AI, and standard AI exclusion wording became available in January. Seven of thirteen legal-malpractice insurers report more AI claims, mostly hallucinations. One frontier lab's IPO filing warns that harm from its technology could exceed the insurance industry's total capacity.

The lesson: autonomy extends exactly as far as someone will underwrite it, which makes the underwriter, not the statute, the effective regulator.

"The underwriter, not the statute, is becoming the real regulator of AI agents."
Artificial Intelligence
Open Weights

A 17B-Active Model Tops the Reasoning Charts

Inkling 2 Small is a 594-billion-parameter multimodal mixture-of-experts that activates just 17 billion per token. It scores 81.7% on ARC-AGI-2 at $0.27 a task and 96.5% on ARC-AGI-1 at eight cents, ahead of every open-weight model tested. Apache-2.0 weights are promised next week.

Its makers also admitted that early checkpoints got around their own guardrails. Kimi K3.1 Pro (2.8T/104B active) and a 1M-context ERNIE Flash, whose licence excludes the EU, UK and South Korea, show open models now splitting by jurisdiction.

Architecture

Inside the Looped Transformer

A looped Transformer runs the same blocks several times instead of stacking new ones. Six unique blocks run four times give 24 block applications, the same depth as a 24-layer model with a quarter of the weights, though without layer-by-layer specialisation.

The key ideas are input injection, adaptive depth (stop looping once the answer is stable) and full- versus middle-stack looping. The claimed payoff is a smaller KV cache, so one GPU can serve more users at once, which matters when memory is the scarcest part of the stack.

Capital

The $80 Billion Question

Anthropic is running at $80 billion a year against a $100 billion year-end target. Investors are debating whether that justifies $1.5 trillion or closer to $1 trillion, and prediction-market odds on the IPO fell 18% while the Nasdaq slid about 10%.

At the same time, forecasters estimate AI is nearly doubling the pace of capability progress itself, with one prominent multiplier revised from 4.5× to 3.8×. The market is effectively pricing the difference between "fast" and "slightly less fast."

Agents & the Engineering Craft
Governance

Agent Policy Lives in Four Places — and None Sees Every Call

One cloud enforces agent tool policy at a single gateway. Microsoft's stack spreads it across four planes: Entra identity, content-safety shields, connector data-loss prevention and framework middleware. None of them answers the one question that matters: may this agent call this tool, with these arguments, right now? Argument-level policy is still custom code.

The details add up. Prompt shields block only in "annotate and block" mode. Tool-response screening for Foundry agents is in preview. Spotlighting is opt-in and costs tokens. Conditional Access for agents requires Agent 365 plus top-tier licences. MCP approvals are opt-in per tool, and Copilot Studio can block only whole servers.

Audit logs default to 180 days, advanced hunting covers 30, and nothing is immutable by default. Agent Framework 1.0 went GA in April and AutoGen is in maintenance mode, so the platform is settled even if the policy layer isn't. The fix is a single enforcement plane in front of every tool call.

Evals

Check the noise band before you roll back. A 2.5-point drop after a harness swap can be pure chance. A three-line binomial calculation tells you whether the regression is real.

Context

Turn your repo into a graph. Index functions, calls and imports, then let agents query them over MCP. The claim is 30–99.7% fewer tokens for questions like "what breaks if I change verify_token?"

Design

A bid can't mark its own homework. Procurement tooling now separates the reader of supplier documents from the scorer, so instructions hidden in a bid can't steer its score.

Hygiene

"Don't ask again" can save your password. Approving a command permanently stores the exact string, credentials included. Keep secrets out of commands and review saved permission rules.

Business & Markets
Infrastructure

A Town That Zones Its Own Supercomputer

Texas elected a governor who may slow, but probably not stop, a rocket company's plan to build compute across the state. That plan includes a multigigawatt "Starbase supercomputer", up to 1 GW of floating power-plant capacity at the Port of Brownsville, 700–800 MW in Bastrop, and data halls inside a chip foundry.

Starbase sets its own zoning and had Brownsville give up land so it could annex it. A 6 GW, 747-acre site in Arizona is the fallback. The company's AI-compute unit nearly tripled in a quarter to $7.56B in revenue, and Musk talks of 10 GW of data centers by 2027. Power has become a question of permits.

Devices

The Memory Squeeze

iPhone revenue rose 17% to $57.6B and Mac 24%, while services hit a record $31.6B but slowed to 10% growth. The new CEO calls memory pricing "unprecedented"; margins are guided down to 45–46%.

In commerce, shoppers arriving via AI assistants convert 64% better than search traffic, and one ad-tech platform runs an 84% EBITDA margin.

Venture

AI Takes 82.7% of the Money

Q3 venture funding reached $515.8B across about 14,200 deals, with AI taking 82.7% of the dollars. Median seed slipped to $3.0M while median Series A rose to $17.4M.

A cleaned list of 894 US investors found that all 191 "family offices" in public databases were really wealth managers. Elsewhere, Stripe bought OpenRouter for $7B, and a small public container runs for $7 a month on Lightsail versus $35 on Fargate.

The Ideas Page — Synthesis & Opinion
Thesis I

Insurance Is the New AI Regulator

Autonomy grows as far as someone will underwrite it. With mainstream carriers excluding AI and specialists pricing it, whoever owns the risk data decides how much autonomy agents get. The opening is an "actuarial trace": a signed, tamper-evident log of agent actions that insurers can price per action, the way usage-based car insurance prices per mile.

Thesis II

The Reader Must Never Be the Judge

Procurement agents that separate reading from scoring, and permission prompts that quietly store secrets, both point to one rule: any component that takes in untrusted content should have no authority over the decision. Build agents with two keys: one to read, one to decide.

Thesis III

Memory, Not Compute, Is the Constraint

DRAM costs take three points off a phone maker's margin, looped Transformers are sold on smaller caches, agents get "dreaming" memory consolidation, and databases ship agent memory. In silicon and in software, the scarce resource is what you can hold, not what you can compute.

Thesis IV

Power Is Permission

Incoming governors want to pause data-center construction while a rocket town annexes land to write its own zoning. Compute is becoming municipal engineering: whoever controls zoning, interconnects and turbines decides where intelligence gets built.

Move 37 · The Contrarian Bet

Bill Agents by the Memory They Keep, Not the Tokens They Burn

Everyone prices AI by tokens, yet memory is what's scarce. Invert it: meter agent services by retained context per day (bytes held hot in caches, vector stores and memory tiers), and pay a "consolidation dividend" to agents that compress their memory without losing task success. Customers pay for the real bottleneck, and vendors are rewarded for forgetting well. The defensible core is a method that attributes multi-tier memory use to individual agent sessions and reprices them in real time based on consolidation quality. Sold to agent platforms on a share of the savings, it has a clear three-year path while memory prices stay high.

— The Daily Signal —