Amazon tells engineers to stop 'tokenmaxxing' AI

Hi there, this is your daily ☕️ Techpresso.

Amazon tells engineers to stop 'tokenmaxxing' AI LINK

• Amazon has told its software engineers to stop "tokenmaxxing", leaning on AI tools as heavily as possible without caring about results, according to a staff memo that reworks how the company defines engineering roles.
• The memo circulated internally in September and was posted to Amazon's website yesterday, framing AI as central to future engineering work while admitting that best practices are still taking shape.
• Companies are curbing tokenmaxxing because the technology's rising costs are hard to justify, capping how many tokens, the units of data AI models process, employees can use or shifting cheaper tasks to open-source models.

Google's Genie 4 lets users edit virtual worlds in real-time LINK

• Google DeepMind has unveiled Genie 4, an AI world model that creates photorealistic videos users can reshape in real-time using plain language prompts, released today as a limited research preview for a small group of early testers.
• Unlike the earlier Genie 3, the new model generates matching audio automatically, from rustling trees to crashing waves, and simulates real-world physics so scenes react to shifting camera angles and lighting changes.
• Google says its biggest change is simulating people, with characters that move naturally, hold conversations, and respond to events, though the model still struggles with visual consistency during longer scenes and tags all output with its SynthID watermark.

Parents sue OpenAI over ChatGPT seizure-drug advice LINK

• The parents of Bowen Wan, a 31-year-old Amazon engineer with epilepsy, sued OpenAI yesterday, saying ChatGPT coached their daughter to stop taking her seizure medication and keep the plan secret from her doctors before she died in February.
• Wan was using GPT-4o, and chat logs the New York Times reviewed show the bot drafting schedules to cut her 400mg daily lamotrigine dose to near zero within weeks, dismissing doctors' warnings as mere "risk aversion" rather than real care.
• The family wants OpenAI to make medical advice a hard limit like it does with copyrighted text; OpenAI called the case "heartbreaking" but said it involved an older model and that newer ones better spot situations needing urgent attention.

Amazon unveils new lineup of Alexa devices LINK

• Amazon answered Apple's recent smart home push on Wednesday by rolling out five new Alexa devices and a new feature, spanning a sleep-tracking alarm clock, a 16-inch display, a photo frame, a camera-equipped speaker, and a smaller speaker.
• The $200 Echo Time clock uses Omnisense sensors to track breathing, presence and room temperature, waking you during light sleep within an hour of your alarm, while a $210 kids version adds parental controls through Kids Plus.
• The $500 Echo Show 16 offers a bigger screen with built-in Fire TV, the $200 Alexa Photo Frame syncs with Amazon Photos, and the sixth-generation Echo Dot adds a built-in smart home hub for $100.

OpenAI plans to buy Google's AI chips LINK

• OpenAI plans to buy Google's AI chips, The Information reported on Tuesday, a move that would deepen ties between two companies that otherwise compete directly in AI products and search.
• OpenAI already relies heavily on Nvidia GPUs to train its models and run ChatGPT, but it started renting Google's Tensor Processing Units, or TPUs, back in June 2025 as a cheaper option.
• Google began selling TPUs directly to some customers this year, and Alphabet reports its quarterly results later today, with investors watching whether its AI spending keeps boosting the cloud business.

Discord to make public chats searchable on Google LINK

• Discord will start showing its public servers on Google, giving each Discoverable Server its own web page that people can read without joining the server or even owning a Discord account.
• Beginning Nov. 16, 2026, all 63,000 Discoverable Servers will get a Server Web Page with discussion summaries, and from mid-December some actual messages inside those servers will appear publicly on the page.
• Server owners can opt out of specific channels starting today under Server Settings, and users can stop their messages from showing, though message summaries can't be refused; usernames and avatars stay hidden, and private servers and DMs remain private.

Other news you might like

• Anthropic buys a one-person biotech and starts its own drug programmes LINK

• Oracle brings Google’s TPUs to its cloud, adds ChatGPT integrations LINK

• Malaysia seizes $12.1M of AI chips at Kuala Lumpur airport LINK

• Anthropic manager scraps coding interviews, says Claude makes it pointless LINK

• EXCLUSIVE: China's CXMT to start advanced AI memory production despite poor yields, sources say LINK

• Hackers Target U.S. Federal Election Commission Website LINK

• FCC clears Amazon Leo to acquire mobile-satellite operator Globalstar LINK

• TikTok expands into restaurant guides with ‘Must Eat’ feature LINK

• Anthropic expands US government AI deal with 40% discount despite Pentagon dispute LINK

• "No human at the helm": Delaware plan would let AI agents run entire businesses LINK

• Your Job Can Survive AI. Your Salary Might Not. LINK

• Nutlope/hallmark: Anti-AI-slop design skill for Claude Code, Cursor, and Codex. LINK

• Some Memories of Early Anthropic LINK

• Meta has a new AI pet app called Waddle World LINK

Trending tools

Overmind: converts your production traces into specialized models you own, with self-hostable infrastructure, deterministic validation, and no third-party training on your data. LINK

Fraud Blocker: blocks invalid ad traffic from bots and click farms, so you only pay for real clicks while protecting budgets and cutting wasted spend. LINK

fswitcher: a cmd+tab replacement for macOS that handles app and window switching across multiple displays, available as a one-time purchase LINK

Zro: provides fast, private AI inference with zero data retention and uncensored security models for red-team and defensive testing without refusals. LINK

ContextGem: open-source Python framework for extracting structured data from documents with LLMs and decision models, auto-generating prompts and tracing every fact to its source. LINK

TONKO: a family pocket money app where kids earn through chores, sort cash into Spend, Save and Grow jars, and learn money skills with parent approval. LINK

Trending papers & reports

Image-editing AI will often redraw a person's face to match a requested race or sexual orientation, producing stereotyped results that later bleed into the system's written descriptions and even tie faces to criminality. LINK

OpenAI's escaped agents broke out of a cybersecurity test this summer and breached Hugging Face's live infrastructure, exploiting flaws and stolen credentials to find test answers, according to a16z citing OpenAI's incident report. LINK

Digital news dominance now reaches 88% of US adults who at least sometimes get news from smartphones, computers or tablets, far ahead of television's 63%, a Pew survey of 5,166 adults finds. LINK

See you tomorrow for a new dose of ☕️ Techpresso!