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

Weekly Fascinating News Briefing

Covering 5 to 12 November 2026 · Macro, AI, agents, crypto, Bittensor, robots and new tech

Strongest conversation starters

Macro Economics

2 stories

Hotter inflation makes a December rate rise look certain

Watch in 30 seconds

US consumer prices rose 0.4% in October, against forecasts of 0.3%. Core prices, which leave out food and energy, rose 0.3% against 0.2% expected. It was the second month in a row that inflation came in hotter than forecast. No yearly figure was published, because there was no October 2025 report during last year's government shutdown.

Fed Governor Christopher Waller had said the night before that a hot reading could justify a rise of 0.5 percentage points at the 8 to 9 December meeting. The 2-year Treasury yield jumped 18.44 basis points to 4.731%, its biggest one-day jump since April 2025, and the 10-year reached 4.988%. Gold fell 2.62% to $4,030.56 an ounce. Brent crude fell to $91.49, its lowest close since 31 August.

October price rises, forecast vs actual (to scale)

Headline, forecast0.3%
Headline, actual0.4%
Core, forecast0.2%
Core, actual0.3%

What traders priced for 9 December (CME FedWatch, per CoinDesk)

Rise of 0.25 points (teal) vs 0.5 points (gold)85% and 15%

Chance of no change: 0%.

Why it mattersHigher rates for longer weigh on anything priced on future growth, from tech shares to bitcoin. One strategist told Reuters the Fed's new worry is AI building work pushing up the cost of "plumbers and electricians".

Nvidia's board is deciding whether to remove Jensen Huang

Watch in 30 seconds

On Sunday 8 November Huang told staff he had an affair with Alice Huang, a former executive at Nvidia customer Megaspeed whom US intelligence reportedly believes is a Chinese spy. The board opened an independent investigation and hired law firm Wachtell Lipton. It said it had no reason to believe national security or company secrets were compromised.

On Wednesday The New York Times reported a second relationship, with Kimberly Powell, Nvidia's vice president of health care, who reported directly to him. The Wall Street Journal then reported the board was preparing to force him out, and the shares fell more than 9% after hours. The board met into the night, but no decision had been announced by Thursday morning. Nvidia reports results on 17 November.

A week of turmoil

8 NovemberHuang admits the affair, board opens a probe
10 NovemberWachtell Lipton hired, shares down 6% on the day
11 NovemberSecond relationship reported, board meets on his future
17 NovemberQuarterly results due
~$600bn
Market value lost since the probe began, if the after-hours fall holds
88%
Polymarket odds he goes by year end
3.6%
Huang's stake in Nvidia (about)
CaveatThe spy claims come from press reports citing US intelligence, and the board's investigation is not finished.
Why it mattersNvidia is the core of the AI trade and has no named successor plan. AMD and Broadcom rose on the news while CoreWeave and SK Hynix fell, so the outcome will move far more than one stock.

AI

3 stories

Anthropic priced its IPO at up to $1.404 trillion

Watch in 30 seconds

Anthropic's updated filing on 10 November offers 825 million shares at $70 to $78 each, raising $57.75bn to $64.35bn. At the top of the range it would be worth $1.404 trillion, up from $965bn in its May fundraise but below the $2 trillion it was reportedly aiming for. AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC will buy another $26bn of non-voting shares at the IPO price, with Google and Nvidia putting in $10bn each.

It expects to start trading on Nasdaq close to 19 November under the ticker ANT, with up to 20% of shares for retail investors. Run-rate revenue hit $81bn at the end of October. Bloomberg reports the deal is already several times oversubscribed.

Anthropic's valuation (to scale)

May 2026 fundraise$965bn
IPO, top of range$1.404tn
Earlier reported target$2tn
$17bn
Third-quarter revenue
$104bn
Net loss over nine months
$900.9bn
Compute and infrastructure commitments
CaveatThe final price is not set yet. Founders keep 50.1% of the votes, so ordinary shareholders will have little say.
Why it mattersThis is set to be one of the biggest listings ever, priced in a shaky market. A strong debut would steady the AI trade; a weak one would hit every AI valuation.

OpenAI is now valued at $1.25 trillion, but raised less than it wanted

Watch in 30 seconds

Bloomberg reports OpenAI raised $11bn at a $1.25 trillion valuation, up from $852bn in March. It had hoped for as much as $30bn at $1.4 trillion before tech shares fell. The Wall Street Journal reports its revenue grew about 50% to $10bn from the second to the third quarter, with an adjusted operating loss of $10.9bn and about 100 million paying ChatGPT users. President Greg Brockman says it could be ready to go public in 2027.

Third-quarter revenue (to scale)

Anthropic$16.9bn
OpenAI$10bn
CaveatOpenAI's figures are from press reports of what it told investors, not a public filing.
Why it mattersAnthropic now earns more per quarter than the company that started the boom, and gets to the public market first.

AI agents proved Signal's encryption works in the real code

Watch in 30 seconds

Google DeepMind used a new system, AlphaProof Code, to mathematically prove key security properties of the Signal messaging protocol, checked right down to the compiled machine code of its main library. Humans approved what had to be proved, then agents running Gemini 4 Carbon wrote about 6.5 million lines of proof, all checked by the Lean proof software. A team of five did it in four weeks, work that DeepMind says normally takes expert teams years.

The job in numbers

6.5m
Lines of checked proof
~140,000
Machine instructions covered
4 weeks
Team of five people
CaveatIt covers the core library, not the whole Signal app, and some low-level parts are assumed rather than proved. One of the two proofs only covers non-quantum attackers.
Why it mattersIf AI can prove code is secure, not just test it, the same approach could check firmware in medical devices and power grids, which DeepMind says is next.

New AI Models

2 stories

Google opened Gemini 4 Carbon, its strongest model, to paying users

Watch in 30 seconds

On 10 November Google made Gemini 4 Carbon available to all paying developers and businesses, and to Google AI Ultra subscribers in the Gemini app and Search. Until now only selected cyber security teams had it. It is built for long, hard tasks like complex debugging and research, and uses more tokens than Gemini 4 Argon to check its work. Google says Argon is still the better choice for most everyday coding.

Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench, at under 50% of the cost per task of the next best outside model. It also previewed /swarm in its Antigravity tool, where one prompt starts a team of agents that plan, build and test in parallel.

Perplexity's WANDR research test, cost per task (to scale)

Gemini 4 Carbon, score 0.857$5.86
Claude Fable 5.2, score 0.851$14.51

API price per million tokens

$2
Input, until 30 November
$10
Output, until 30 November
$4
Input, after that
$20
Output, after that
CaveatThe rankings are partner benchmarks quoted by Google. On a new test Google ran with Artificial Analysis, Carbon came second, with 54% against 57% for Claude Fable 5.2.
Why it mattersCarbon's price doubles after 30 November. Teams that pair one expensive "thinker" with cheaper models for routine work, as /swarm allows, will spend far less.

AlphaGenome 2 predicts how DNA changes affect individual cell types

Watch in 30 seconds

Google DeepMind released AlphaGenome 2 on 11 November. The first version predicted how DNA changes affect broad tissues. The new one works cell type by cell type, such as a heart muscle cell or an immune cell, and can handle new cell states from a sample of just a few dozen cells without retraining. It was trained on more than 107 million single-cell profiles across 68 tissues. In one test it reproduced a known finding: the DNA change behind a form of inflammatory bowel disease, acting only in gut immune cells.

Immune cell test: share of causal DNA changes found (to scale)

Original AlphaGenome, fine-tuned16%
AlphaGenome 230%
Direction of gene effect, original23%
Direction of gene effect, AlphaGenome 243%
CaveatThese are DeepMind's own results. It is for research only, free for non-commercial use, and not approved for diagnosis.
Why it mattersMost disease-linked DNA changes sit in switches that only matter in certain cells. Pinpointing which cell is affected is what drug hunters need.

AI Agents: Capability, Reliability and Safety

3 stories

An AI attack agent called ARTEX is now being used against real companies

Watch in 30 seconds

Blockchain infrastructure firm Chainstack said on 10 November that an attacker using the ARTEX AI agent framework broke into its systems in late October. The way in was a Grafana dashboard on a test server, open to the internet and still on its default admin password. From there the attacker took machine credentials that reached production and the database network. Customer names, emails and password hashes were exposed. Chainstack says nodes, console and API were not changed, but it is treating node credentials as exposed and asking customers to rotate them.

Separately, Microsoft said on 6 November that a Cambodia-based criminal group it calls Storm-3358 began using ARTEX against financial firms in October. The agents scan for weak spots, exploit them and install remote access tools "at machine speed".

How the Chainstack breach unfolded

26 Oct: InTest dashboard with its default password
29 Oct: SpreadMachine credentials reach production
ExposedNames, emails, password hashes
31 Oct: Cut offSpotted and access paths closed
CaveatThe account of the breach comes from Chainstack's own blog post.
Why it mattersAI agents make old mistakes, like default passwords on forgotten test servers, far quicker to find and exploit. Microsoft's advice includes decoys and fake "wins" that waste an attacking agent's time.

A flaw in an AI engine let a hacker into NEAR AI's agents

Watch in 30 seconds

On Sunday 8 November an attacker used a flaw in SGLang, a popular open-source engine for running AI models, to run code on one of NEAR AI's servers. Its "confidential mode" runs requests in a sealed chip area, but that does not protect against bugs in the software running inside it. The attacker took chat history, settings and logs from 18 older AI agents, account details for about 100 users and two customers' API keys, which have been revoked.

18
Agents accessed
~100
Users' details taken
2
API keys exposed
CaveatNEAR AI has not named the affected SGLang versions or said whether the project has released a fix.
Why it mattersPrivacy hardware only helps if the software inside it is safe. Anyone running SGLang should check for updates.

ChatGPT now lets brands put their own agents in front of shoppers

Watch in 30 seconds

OpenAI launched a beta of "Business Agents" on 10 November, first for eligible US online shops and services. When a business shows up in a ChatGPT answer, you can chat with its agent about products and ask follow-ups. The chats are labelled and kept separate from your main conversation. Firms claim and verify their profile through a new Business Hub, which also shows how often ChatGPT mentions them.

CaveatOpenAI has not said exactly how it verifies businesses.
Why it mattersThis makes ChatGPT more like a search engine with a shop window. Being visible in AI answers is becoming as important as ranking on Google.

Crypto

3 stories

Bitcoin fell below $73,000, its weakest since August

Watch in 30 seconds

After the hot inflation report on 10 November, bitcoin dropped below $73,000 for the first time since 20 August, down 3% in 24 hours, while ether, XRP and solana fell closer to 6%. It was also five years to the day since bitcoin first hit $69,000. Selling has come from companies too: according to Lookonchain, 14 listed companies have sold a combined 108,247 BTC this year, about $8.08bn, led by miners MARA and Riot.

Bitcoin sold by listed companies in 2026 (to scale)

All 14 companies108,247 BTC
MARA40,581 BTC
Riot Platforms20,264 BTC
CaveatThe company sales figures are from on-chain tracker Lookonchain, as reported by CoinDesk.
Why it mattersRate rises, a wobbling Nvidia and miners selling coins are all hitting at once. CoinDesk notes the pattern looks like the spring, when a rally faded back to $60,000.

MSCI is dropping Strategy from its stock indexes

Watch in 30 seconds

MSCI confirmed it will remove Strategy, the biggest listed bitcoin holder with 850,364 BTC, from its global indexes after the close on 30 November, under new rules for "non-operating" companies. Tokyo's Metaplanet is also out. Funds tracking MSCI hold about 13 million Strategy shares, 3.1% of the total and worth about $1.6bn, which they will have to sell.

Share price vs value of its bitcoin, mNAV (to scale)

Peak, November 20243.4x
Now~1.08x
CaveatSeveral analysts say the move was widely expected and is mostly priced in.
Why it mattersWith little premium left, selling new shares to buy bitcoin barely adds value. In its last week Strategy spent more buying back preferred stock than buying bitcoin.

Coinbase won a full US national trust charter

Watch in 30 seconds

The US banking regulator, the OCC, upgraded Coinbase's trust from conditional to a full charter on 9 November, after Ripple got the same last month. It lets Coinbase offer federally regulated custody and payments, but not take deposits or make loans like a full bank. A community banks group is suing the OCC over these charters.

Why it mattersCrypto firms are moving inside the US banking system, which makes it easier for big institutions to use them for custody and stablecoins.

Bittensor

3 stories

Gamma is live: pay for subnet services from a wallet, and burn alpha doing it

Watch in 30 seconds

The Opentensor Foundation formally launched Gamma on 11 November, after it was switched on the night before. Gamma is a dollar-valued prepaid credit. Users and AI agents can pay for subnet services such as inference, compute and storage straight from a Bittensor wallet, without a separate account or API key for each subnet. Requests are signed with a wallet hotkey instead.

Credits can be bought with TAO, which buys and burns the subnet's alpha, or by burning alpha already staked. Alpha worth $100 makes 100 credits. Credits cannot be cashed out, moved to another coldkey or spent on a different subnet, and spending them destroys them rather than paying the provider in tokens.

How Gamma works

1. FundPay with TAO or staked alpha
2. BurnThat subnet's alpha is burned
3. Credit$100 of alpha makes 100 credits
4. UseSign requests with a hotkey, credits spent
CaveatEach subnet and provider must opt in, and Taostats was still asking which ones had. Robotensor says SN58 (Attune) is the first to enable it, which has not been checked independently.
Investor takeawayThis ties paying customers directly to burning subnet alpha. If real demand shows up, usage starts to shrink token supply.

Quantum-resistant wallets are on mainnet, but not switched on yet

Watch in 30 seconds

Runtime 488 went live on 5 November at block 9,442,681 and adds three post-quantum signing schemes: ML-DSA-65, Falcon-512 and Falcon-1024. It also adds optional key rotation, where each transaction moves the account to a new signing key while keeping the same address, balance and recovery phrase. Existing wallets keep working and nobody has to migrate.

The new accounts cannot sign until root switches the feature on. Post-quantum addresses can already receive TAO, but cannot spend it yet.

Signature size in bytes (to scale)

Falcon-512, lowest security level666
Falcon-1024, highest security level1,280
ML-DSA-65, NIST standard3,309
CaveatDo not move funds to a post-quantum address until activation is confirmed and your wallet software supports it. Ethereum-style EVM wallets on Bittensor are not covered.
Investor takeawayBittensor now has a path to quantum-safe accounts, but holders are only protected once the feature is switched on and they move to a new account.

Root basket deposits have been stuck since 7 November

Watch in 30 seconds

Since 7 November the team has confirmed that root dividends are building up but cannot enter stakers' baskets. On 11 November a community member said a fix was in the release branch and guessed it would ship the next day, but no official upgrade announcement had appeared. Among the week's merged changes is a fix for hotkeys stuck with more than 256 pending basket deposits.

On the releases page, runtime 489 is a proposed upgrade still waiting for the second approval from the triumvirate multisig, and runtime 490, tagged on 10 November, only moves the version number on.

The week on chain

5 NovemberRuntime 488 live with Gamma and quantum wallets
7 NovemberBasket deposit block confirmed
10 NovemberRuntime 489 proposed, 490 tagged
PendingOfficial fix announcement
CaveatThe "fix is in the release branch" claim came from a community member, not the team. Bittensor Daily did not confirm that the 256-entry fix is the same issue.
Investor takeawayIf you stake on root, check whether your basket balance is updating before making changes.

Robots

3 stories

Nvidia's GR00T 2 learns robot jobs from videos of people

Watch in 30 seconds

Nvidia unveiled GR00T 2 on 9 November at the CoRL robotics conference. It predicts future video and robot actions together, so it learns how movements change the world around it. It was trained on 402,730 hours of first-person human video plus robot demonstrations, and Nvidia is releasing 100,000 hours of that human video for commercial use. Foxconn is applying it to server manufacturing, and code is due in December.

Results across 50 handling tasks (to scale)

Tasks done at least once with no robot demos80%
Average success, 1 demo per task71%
Average success, 5 demos per task95%

80% is 40 of the 50 tasks. Five demos took under 3 minutes of robot data and about 30 minutes of training on one RTX 5090.

CaveatNvidia's paper says the 50 tasks were picked from ones that already reached at least 80% success with five demos, so this is not proof it works on any new task.
Why it mattersCollecting robot demonstrations is slow and costly. Learning from ordinary video of people could cut the time needed to teach a robot a new task.

Google's new robot brain hit 96% in a Mercedes-Benz factory trial

Watch in 30 seconds

Google DeepMind introduced Gemini Robotics 2.5 on 11 November. In a trial on Mercedes-Benz production lines in Hungary and Germany, Apptronik's Apollo 2 humanoid picked metal brackets from bulk bins and placed them in order for each car, with a reported 96% success rate. In lab tests the model beat earlier versions by 30% to 100% on success rate, and was 20% to 30% faster.

96%
Success in the car-parts trial
200+
Tasks tested
43%
More safe completions with people nearby
CaveatThe 96% applies to this one trial. The model runs on cloud servers, and it is only going to Google's robotics partners.
Why it mattersCar plants are a key early market for humanoids, and sorting parts for assembly is a real job, not a demo.

Runway gave away its robot control model for anyone to use

Watch in 30 seconds

On 11 November Runway, best known for AI video, released Praxis-1 with open weights under the Apache 2.0 licence, including for commercial use. It is a 27 billion parameter model that predicts video and robot actions together, and runs on anything from data centre chips to the Jetson Thor computer inside a robot.

RoboLab-120 success rate (to scale)

Praxis-146.2%
FLUX 3 Action42.9%
Cosmos3 Nano36.8%
π0.528.0%
CaveatScores are as published by Runway. Even the best model fails more often than it succeeds on this test.
Why it mattersFree, adaptable robot brains lower the cost for small robot makers who cannot train their own.

New Technology

3 stories

British AI chip startup Fractile raised $720m

Watch in 30 seconds

Fractile announced a $720m Series C on 11 November, led by Thrive Capital, Founders Fund, Lightspeed and Redpoint and backed by the UK's Sovereign AI Fund. The UK government put the valuation at over £5bn. Its chips stack ordinary DRAM memory directly on top of each chip, to run large AI models fast without the usual trade-off between memory and speed. It plans to ship its first chips and systems next year.

$720m
Series C
£5bn+
Valuation, per the UK government
2027
First chips due
Not demonstratedNo chips have shipped yet, so the speed claims are untested.
Why it mattersPurpose-built chips are one way to cut the cost of running AI models. A UK-backed chip maker gives Britain a stake in that hardware.

A silicon-chip quantum computer went into a normal data centre

Watch in 30 seconds

Diraq installed what it calls the first silicon spin quantum computer inside a commercial data centre, at Equinix in Sydney, announced on 9 November. It is made with the same processes as ordinary computer chips. It runs with 8 qubits today, and Diraq is aiming for up to 256 in the same system within a year.

8
Qubits now (256 targeted)
<20 kW
Power draw
170 sq ft
Floor space
CaveatThis is a company press release. Eight qubits is far too few for useful work.
Why it mattersIf quantum computers can be built on normal chip lines and fit in normal data centres, they get much cheaper to scale.

Anduril's satellite can track other satellites in Earth's shadow

Watch in 30 seconds

Anduril said on 10 November its first satellite, Anduril-216, has shown it can spot other spacecraft using heat-sensing infrared, even in full darkness. Flying about 590 km up, it spends about 33% of each orbit in Earth's shadow, where normal optical tracking goes blind and satellites can move unseen. Anduril plans to sell the tracking data to the US military.

CaveatAnduril has released only one image so far, of the Moon.
Why it mattersKnowing when a rival satellite moves is a growing military concern, and private firms are starting to fill the gap.