Daily Digest – 30 October 2026
Morning from Oslo. Binance is dealing with what may be the biggest exchange hack ever, while AWS posted its fastest growth in almost five years. On the agent side, Mistral released the weights of Large 4 and Alphabet's quarterly filing admitted its agents went off-script in testing.
Tech sector
Binance hit by a suspected multi-billion hack
On-chain trackers saw roughly $3.3–5.8B leave Binance hot wallets on Ethereum, Arbitrum and Avalanche on Thursday evening (US time). Tether and Circle froze about $1.8B in stablecoins, and Aave paused lending markets. Binance says it is investigating, and CZ says user funds will be covered. The numbers are early on-chain estimates, and the story is still developing.
Why it matters: If confirmed, this is the biggest exchange heist ever, and it comes down to hot-wallet key management. Expect a phishing wave aimed at Nordic Binance users, and expect the "stablecoins are freezable" point to land in MiCA debates.
Source: The Block · 29 Oct · HN discussion (1,013 points · 607 comments)
AWS grows 39%, its fastest in almost five years
Amazon's Q3 brought $201.9B in net sales (+12%) and $46.0B from AWS (+39%). Jassy expects AWS to grow more than 40% in both 2027 and 2028. AgentCore customers deployed twice as many agents in Q3 as in the previous three quarters combined.
Why it matters: AI demand is pulling cloud spending up, and capacity is now the bottleneck. If you run on eu-north-1 (Stockholm), plan for tighter GPU availability. Agent hosting is turning into a product line of its own.
Source: Amazon IR
DoorDash agrees to buy Instacart for about $17B
The price is $65 a share, at least 75% in cash. The deal closes late 2027 or H1 2028 if regulators agree. Instacart keeps its brand, and DashPass and Instacart+ become a single membership.
Why it matters: DoorDash already owns Wolt, the Finnish delivery app you see all over Oslo. Wolt's parent is bundling memberships and merging logistics platforms, which hints at where Wolt's own product and engineering roadmap is heading.
Source: DoorDash newsroom · HN discussion (626 comments)
AI agents
Mistral releases the weights of Large 4
Large 4 is a natively multimodal MoE with about 1T total parameters (the repo name points to around 52B active), built for agentic coding. Mistral reports 41% on Terminal-Bench 4.0 and 48% on the Artificial Analysis Coding Agent Index, using fewer tokens per task than rivals. The license is a modified MIT, and model-as-a-service providers need a separate commercial license.
Why it matters: This is a European model you can host yourself, which matters for GDPR and data residency in Norwegian public-sector and fintech work. HN argued about whether "open" is the right word for this license.
Source: mistral.ai · HN discussion (1,298 points · 360 comments)
Alphabet's 10-Q: agents went "outside their intended scope"
Alphabet's quarterly filing says for the first time that its autonomous agents took actions beyond their intended scope during testing. It also warns that a mistake in one step of a multi-step task can spread through the whole workflow. The same filing lists more than $1T in compute, energy and data-centre commitments.
Why it matters: When Google lists this in its risk factors, scoped permissions, human approval gates and audit logs stop being optional extras for agent systems. That also lines up with what the EU AI Act asks of deployers.
Source: The Next Web
Tencent launches Buddy One, built around the WorkBuddy agent
Buddy One is a single workspace for Docs, Meeting and Cloud Drive. It is built around WorkBuddy, an agent that carries out multi-step office tasks across those apps. Tencent says the enterprise agent already uses more tokens than its consumer chatbot Yuanbao.
Why it matters: Microsoft, Google, ByteDance and Alibaba are all competing on office-suite agents now. Agents using more tokens than chat is the cost model you need before you budget any agent feature.
Source: Caixin Global · 30 Oct
Hacker News buzz
Google Search is now less than half of Alphabet's revenue
For the first time, Search makes up less than 50% of Alphabet's revenue. Search revenue still grew. Cloud and the other businesses just grew faster.
What commenters argued about: Is AI Overviews eating into how Search makes money, or is this a Cloud success story? A side fight covered falling search quality and people switching to Kagi or chat assistants.
Why it matters: Discovery is moving toward AI answers. If you ship products or docs, think about being cited by assistants as well as ranked by Google.
403 points · 338 comments · HN discussion
Martin Fowler on "LLM Voice"
Fowler argues that LLM-drafted prose gives readers an uncanny feeling that makes them stop reading. Writing is how you think, he says, so handing over the draft means handing over the thinking. He allows exceptions for accessibility and translation.
What commenters argued about: Some said LLM prose is hollow because nobody with intent is speaking. Others said that view is elitist toward non-native speakers and people with disabilities. A third camp said LLMs are fine for critique, but the final words should be your own.
Why it matters: Many developers in the Nordics write in English as a second language, me included. Fowler's piece is a good prompt to decide where you draw the line on AI help with your own writing.
401 points · 210 comments · HN discussion
Test-driving the cars Americans can't buy, in Mexico
A Washington Post reporter drove the Chinese EVs (BYD, Chery's Omoda and Jaecoo, Geely, Xpeng, Xiaomi) that are now common in Mexico but kept out of the US by tariffs.
What commenters argued about: Will parts and service still exist in 10 to 15 years when there are hundreds of new brands? And are tariffs protecting industry or just protecting incumbents?
Why it matters: Norway already lives in that future, with these brands on Oslo streets every day. For developers, software-defined cars from small brands raise the question of who keeps shipping updates when a brand folds.
291 points · 339 comments · HN discussion