Archived edition · 22 to 29 October 2026Earlier: 15 to 22 OctoberNext: 29 October to 5 NovemberLatest edition: 12 to 19 November →
Siam Kidd · Weekly briefing

Weekly Fascinating News Briefing

Covering 22 to 29 October 2026 · Macro, AI, agents, crypto, Bittensor, robots and new tech

Strongest conversation starters

Macro Economics

3 stories

The Fed held rates, but two policymakers already wanted a hike

On 28 October the Federal Reserve kept its main rate at 3.75% to 4% in a 10 to 2 vote. Governor Lisa Cook and Cleveland Fed President Beth Hammack voted for a 0.25 point increase. The statement said inflation "remains elevated" and described the economy as "expanding at a strong pace".

Chair Kevin Warsh also said one of the eight meetings in 2027 will focus only on long-term strategy, with no rate vote.

How the 12 voters split

10 voted to hold at 3.75% to 4%2 wanted a 0.25 point hike: Lisa Cook and Beth Hammack

The decision in numbers

3.75–4%
Rate range, unchanged
10 to 2
Vote, both dissents for a hike
2%
Inflation goal the Fed says it will deliver
Why it mattersMarkets were watching for a December hike. Two dissents for a rise suggest the next move is more likely up than down.

The UK Budget: energy help, paid for by banks, savers and landlords

Chancellor John Healey's Budget on 28 October gives a £100 energy payment to up to 20 million households in 2027 and extends the 5p fuel duty cut by six months. The bank tax surcharge doubles from 3% to 6% for 15 months, and tax on savings and property income rises by 2 points from April 2027.

The OBR cut its 2027 growth forecast to 1.0%, from 1.6% in March. There is also a new Founders Relief: a 10% capital gains rate on up to £10m of lifetime gains for founders of eligible high-growth companies.

2027 growth forecast (to scale)

OBR, March1.6%
OBR, October1.0%

Who pays more

BanksSaversLandlordsGamblingOnline marketplaces (VAT)
Why it mattersThe safety margin against the fiscal rules has shrunk to £15bn, from £24bn, mostly because of higher debt interest. There is little room left if borrowing costs rise again.

Economists now expect high inflation to last into 2027 in most big economies

A Reuters poll of more than 500 economists, published on 27 October, raised 2027 inflation forecasts for 62% of the 50 economies covered, and for 71% of developed economies. Oil has stayed above $100 a barrel since 7 October and recently topped $117, driven by the war with Iran. Global growth is still forecast at 3.0% this year and 3.2% next year.

Share of economies with higher 2027 inflation forecasts (to scale)

All 50 economies62%
Developed economies71%
Why it mattersGrowth is holding up, but prices are not coming down. That combination keeps interest rates higher for longer, which is the backdrop for every other story on this page.

AI

2 stories

Anthropic filed to go public, and the numbers are huge in both directions

Anthropic's IPO filing on 26 October shows revenue of $16.3bn in the first half of 2026, up about 1,300% on a year earlier. It also shows a net loss of $80.7bn, but $72.4bn of that is a non-cash accounting charge on convertible notes. The operating loss was just $693m, and operating cash flow turned positive at $3.9bn.

It plans to list on Nasdaq as "ANT", reportedly in Thanksgiving week. Media reports put the target valuation at $1.5tn to $2tn. The seven founders keep 50.1% of the votes through a special share class.

First half of 2026

$16.3bn
Revenue, up about 1,300%
$3.9bn
Operating cash flow, first time positive
$2.08of revenue for every $1 spent on compute (Q2)

Revenue against loss, first half of 2026 (bars to scale)

Revenue$16.3bn
Operating cash flow (positive)$3.9bn
Operating loss$693m
Net loss$80.7bn
Non-cash charge on convertible notes, $72.4bnRest of the net loss, $8.3bn

The accounting charge is about 90% of the net loss. The operating loss is less than 1% of it.

Animated explainer: Anthropic revenue against loss
CaveatThe valuation range comes from media reports, not the filing. The filing also flags risks: the top five customers were about 25% of revenue, and nearly half of revenue flows through Amazon, Google and Microsoft cloud marketplaces.
Why it mattersThis is the first full set of audited numbers from a frontier AI lab. It shows the business can generate cash, and it will set the price for every AI listing that follows, including OpenAI.

Meta plans to raise $95bn from shareholders to pay for AI

On 26 October Meta said it would raise $35bn through public share offerings, then up to $60bn more by selling shares gradually on the market. Nvidia, AMD and Broadcom showed interest in buying up to $5bn of the shares. Meta now expects $135bn to $145bn of capital spending this year and is building four data centre "superclusters" of more than 1 gigawatt each.

How the $95bn is split (to scale)

■ Upfront offerings■ Gradual sales on the market
Why it mattersEven the most profitable tech firms can no longer pay for AI from cash flow alone. Issuing shares dilutes current owners, which is a sign of how expensive the AI race has become. The shares fell 1.6% that morning.

New AI Models

3 stories

Google's Gemini 4 Carbon tops the rankings, and goes to cyber defenders first

Announced on 28 October, two weeks after Gemini 4 Argon, Carbon ranks first on the Artificial Analysis Intelligence Index with a score of 59. Google says it leads benchmarks for finding, exploiting and fixing software flaws. Because of that, access starts with a small group of cyber defenders, with developers and consumers to follow.

Security benchmark scores (to scale)

SEC-bench Pro: finding flaws88.2%
CWE-bench: patching flaws71.0%
ExploitGym: exploiting flaws54.8%

Price per 1M tokens, input / output

$2 / $10
Introductory price
$4 / $20
After the intro period
CaveatThe benchmark scores come from Google and were measured without guardrails. Independent testing will follow wider access.
Why it mattersFrontier labs are now treating hacking ability as the reason to restrict a release. Defenders get a head start, but attackers will eventually have similar tools.

OpenAI's GPT-6.1 Terra halves the price of its computer-use model

Released on 26 October, Terra is OpenAI's fast, cheap model for everyday computer tasks such as filling in forms and reading invoices. API prices are 50% lower than GPT-5.6 Terra, at $1 per million input tokens and $5 per million output tokens. On the Terminal-Bench 4.0 coding test it solves 50.3% of tasks, against 21.5% for the previous version.

Error rate on tricky prompts, low effort (to scale)

GPT-5.6 Terra53.6%
GPT-6.1 Terra16.7%
CaveatAll figures are OpenAI's, including its claim of about 9 times lower cost per task than Claude Opus 5.5. OpenAI also says the model reaches its "Critical" cybersecurity risk level.
Why it mattersCheap, reliable computer use is what turns office work like data entry into something agents can do at scale.

Genie 4 generates explorable worlds with sound, in real time

Google DeepMind's Genie 4, announced on 27 October, builds interactive worlds you can walk around and change with text prompts as they run. It is DeepMind's first real-time world model to generate audio too, including character dialogue. Gemini runs alongside it to keep scenes consistent and plan multi-step events.

Who can use it

NowTrusted testers, research preview
SoonGoogle AI Ultra subscribers
SoonGemini API early access partners
CaveatDeepMind says details can still drift over longer sessions, and text, physics and audio accuracy are still being worked on.
Why it mattersWorld models are how robots and agents may be trained in future: cheap, endless simulated practice before touching the real world.

AI Agents: Capability, Reliability and Safety

2 stories

A file in your Downloads folder can take over your AI coding agent

Forever Security found a malicious AGENTS.md file in a customer's Downloads folder that told a coding agent to install a rogue npm package. Agents read these instruction files automatically from the folder they start in, and most browsers save downloads there without asking.

In tests, the trick worked on OpenAI Codex 90% of the time with a harmless payload. A similar CLAUDE.md file redirected Claude Code's traffic and exposed its login token, leading to code execution in about 50% of tests. Claude Code alone had more than 1,300 sessions launched from Downloads folders in the firm's data.

How the attack works

1. DownloadA web page saves AGENTS.md into Downloads
2. LaunchThe user starts a coding agent in Downloads
3. ReadThe agent reads the file as its instructions
4. RunIt installs the attacker's npm package

How often it worked in the researchers' tests

OpenAI Codex, AGENTS.md, harmless payload90%
Claude Code, CLAUDE.md redirect, code ranabout 50%
Animated explainer: how a downloaded AGENTS.md file hijacks a coding agent
CaveatThe researchers could not confirm how the file reached the customer's machine. The success rates are from their own tests.
Why it mattersThere is nothing for OpenAI or Anthropic to patch: the agent is doing exactly what it was built to do. The fix is habit. Always start agents inside the specific project folder, never in Downloads or your home folder.

A hijacked release of LiteLLM, a popular AI gateway, stole credentials

On 21 October a compromised LiteLLM 1.107.0 was published to PyPI and the official container images. Microsoft says it ran as soon as Python started, even if LiteLLM was never used, and collected cloud keys, AI provider keys, SSH keys and browser data. The attacker used stolen developer GitHub credentials to redirect the release process.

What Microsoft says to do

Find every copy of 1.107.0Do not run Python inside a suspect environmentRebuild from clean imagesRotate every reachable credential
Why it mattersAI gateways sit next to the most valuable keys a company has. Agent tooling is now a supply-chain target, and simply removing the package does not clean up the damage.

Crypto

3 stories

Bitcoin slipped below $80,000 ahead of the Fed

On 26 October Bitcoin fell nearly 2% to about $79,800 as traders cut risk before the Fed meeting and AI stocks sold off in Asia. Ether dropped nearly 3% to just above $2,300. Traders said Bitcoin needs to get back above $82,500 to restore confidence.

Levels to watch

~$79.8k
Low on 26 October
$82.5k
Level lost on 22 October
Why it mattersCrypto is trading like a tech stock on rate news. With two Fed officials now voting for a hike, the macro headwind has not gone away.

The US bank regulator wants small banks to issue their own stablecoins

On 26 October the OCC urged community banks to apply to become stablecoin issuers under the GENIUS Act. Its rules take effect on 18 January 2027, although applications have not opened yet. Separately, community bankers are suing the OCC over trust charters it gave crypto firms.

The path to issuance

26 OctoberOCC invites community banks
PendingApplications open after federal sign-off
18 January 2027Federal approval needed to issue
Why it mattersStablecoins could spread through thousands of small local banks, not just big banks and crypto firms.

Canada's draft stablecoin rules ban paying interest to holders

Canada opened a 60-day consultation, ending 23 December, on its draft stablecoin rules. Issuers must hold reserves equal to every token in circulation at Canadian banks, credit unions or trust companies. Holders must be able to cash out within three business days, with fees capped at 1%. Interest is banned, even through third parties, but rewards linked to transactions are allowed. The rules are expected to take effect in late 2027.

The draft rules at a glance

100%
Reserves held in Canada
3 days
Maximum redemption time
1%
Maximum redemption fee
Why it mattersBanks on both sides of the border are winning the argument that yield-paying stablecoins would drain deposits. Expect "no interest" to become the global standard.

Bittensor

2 stories

The first wave of US TAO fund filings has arrived

On 23 October Grayscale filed an amended registration to turn its Bittensor Trust into a spot TAO ETF on NYSE Arca, keeping the GTAO ticker. Shares would be created and redeemed in TAO or cash, which should keep the price close to the value of the TAO held. Bitwise and REX-Osprey have also filed TAO products.

The trust held about $16.2m of assets on 23 October. Its current fee is 2.50% a year, and staking is not allowed until a tax condition is met.

TAO fund filings timeline

23 OctoberGrayscale files Amendment No. 2 to its S-1
By 27 OctoberBitwise and REX-Osprey have also filed TAO products
PendingGrayscale's registration becomes effective
ThenLists on NYSE Arca as Grayscale Bittensor ETF, ticker GTAO
$16.2m
Trust assets on 23 October
2.50%
Current yearly fee
22%
GTAO premium to asset value on 23 October (X post)
Animated explainer: the TAO fund filings so far
CaveatThese are filings, not approvals. The reported 22% premium of GTAO over its asset value comes from a post on X.
Why it mattersAn ETF lets ordinary brokerage and pension accounts buy TAO. The OTC trust's premium suggests demand for regulated exposure is already ahead of supply.

Four runtime upgrades in a week, with quantum-resistant wallets next

Runtimes 483, 484 and 486 went live between 23 and 26 October. Runtime 486 makes builds reproducible, so anyone can check that the published code matches what runs on chain. It also fixes a fee-accounting gap on a stake query. Release 487, reported on 28 October, adds optional quantum-resistant wallet types (ML-DSA-65, Falcon-512 and Falcon-1024) and a narrower proxy that can manage stake locks but cannot move stake.

Last week's configurable emission split is already in use. SN78 (Umi) moved to 30.8% owner, 15.5% validators and 53.8% miners.

SN78 emission split, old default vs new

■ Owner■ Validators■ MinersTop: old default. Bottom: SN78 now
CaveatRelease 487 was not yet listed on the subtensor GitHub releases page when checked on 29 October. Existing wallets keep working, and the new key types are optional.
Investor takeawayCheck each subnet's emission split. SN78 cut validators' share by more than half, which changes who is paid to support its token.

Robots

3 stories

Tesla has started early production of Optimus in California

Elon Musk said on Tesla's earnings call that early Optimus production has begun at Fremont. Most of a line built for 1 million robots a year is installed, but reaching that rate will take "a year, maybe longer". Customer shipments are planned for next year, first for dangerous, repetitive industrial work. Tesla also plans an "Optimus Academy" of 10,000 to 20,000 robots that practise tasks to generate training data.

Humanoids are still a tiny share of robots bought last year (to scale)

All robots added worldwide603,000
Humanoids bought~7,000

Humanoids were about 1.2% of the total, per the International Federation of Robotics.

Not demonstratedThe 1 million a year capacity and delivery timing are Tesla targets. No customer robots have shipped.
Why it mattersIf Tesla gets close to its targets, humanoid volumes would jump far beyond today's market of a few thousand a year.

Google's robot "brain" model is now available to businesses

Gemini Robotics ER 2 went into public preview on Google Cloud on 27 October. It plans tasks, checks from video whether each step worked, and directs the robot's movement software. Google says it works on everything from fixed arms to humanoids, and that it pinpoints the key moment in a video, such as when to stop pouring, with 91.3% accuracy. Boston Dynamics is using it for Spot inspections.

CaveatThe accuracy figure is Google's own measurement.
Why it mattersRobot makers can rent the reasoning layer instead of building it, much as app makers rent cloud computing.

Alibaba taught a humanoid mostly by having it watch people

Alibaba's Lambda-0 model was trained on 500 hours of first-person video of people doing everyday tasks, then fine-tuned on a Unitree G1 robot. It succeeded in 62.5% of real-world trials, against 40% for the best rival model tested. Recording a human demonstration took about a quarter of the labour of a robot demonstration.

Real-world success rate (to scale)

Lambda-062.5%
Best rival (π0.5)40%
CaveatThese are the authors' own lab results from 40 trials, not yet reproduced by others. The human video dataset has not been released.
Why it mattersTraining data is the bottleneck for robots. Cheap human video could be the robot equivalent of the internet text that trained chatbots.

New Technology

2 stories

Apple put a touchscreen on the Mac for the first time

The new MacBook Pro, announced on 28 October, has a tandem OLED screen with up to 2,000 nits brightness, touch built into the display and the iPhone's Dynamic Island. It runs on M5 Pro or M5 Max chips, which Apple says can run AI models with hundreds of billions of parameters on the laptop itself. Apple claims up to 24 hours of battery life. Prices start at $2,899, with deliveries from 6 November.

What is new

TouchscreenTandem OLED2,000 nitsDynamic IslandUp to 24 hours battery (Apple's test)Thinner and lighter
CaveatBattery life and performance figures are Apple's own claims.
Why it mattersApple held out against touch on the Mac for years. Powerful local AI on laptops also reduces reliance on cloud models for some work.

IQM got all 54 qubits on a cloud quantum chip working together

On 29 October Finnish firm IQM said its cloud-accessible Emerald system entangled all 54 qubits on the chip at once. Before this, users could reliably use about 30 qubits in a single program. Calibration work cut error rates in quantum operations by up to 50%.

Entanglement fidelity vs the 0.500 threshold for genuine entanglement

0.516
All 54 qubits, with readout correction
0.515
40 qubits, without correction
0.500
Threshold
CaveatThis is a company announcement, and the result clears the threshold by a narrow margin. It is a benchmark, not a useful calculation.
Why it mattersMost experiments on large quantum chips only use 10 to 20 qubits. Using the whole chip at once is a step towards useful quantum computing.