Overview
Today’s feed moved on two tracks: new AI modes that promise faster voice and sharper reasoning, and a sober mood in markets and work. Microsoft pushed realtime voice to the edge, Google rolled out a beefed-up reasoning mode, and Anthropic opened a window into how people feel about AI at work. Elsewhere, layoffs ticked up, Michael Burry warned of a long selloff, Apple’s AI house politics spilled into public, and debates flared over attention, software licences, and what “tech” should mean.
The big picture
Realtime voice at the edge
Microsoft’s VibeVoice-Realtime-0.5B is a compact text-to-speech model built for streaming with roughly 300ms first sound on modest GPUs like a T4. It posts strong zero-shot scores against VALL-E 2, runs on an English Qwen2.5-0.5B backbone, and ships under MIT with watermarking - a clear nod to developers building voice on devices, though it is single speaker and English-only for now.
Reasoning steps up in consumer apps
Google’s Gemini 3 Deep Think mode has started rolling out to Ultra subscribers. Jeff Dean highlights better generalisation on hard tests like ARC-AGI-2 and strong results on HLE and GPQA Diamond, with a daily cap to manage use. Sundar Pichai points to a 3D architecture demo, and users are already asking for a “project brain” that can reason across Docs and Drive over weeks.
Listening at scale
Anthropic launched Interviewer - a short pilot that drafts questions, runs adaptive chats, and analyses responses with humans in the loop. Early findings from 1,250 professionals suggest big time savings, creatives feeling stigma for AI use despite gains, and scientists keeping AI away from core tasks due to trust gaps.
Apple’s AI reckoning
Mark Gurman says Apple missed its generative AI targets, pushed Siri’s overhaul back by 18 months, and eased its AI chief into retirement with garden leave. Amar Subramanya steps in as Apple tries to quicken AI work across devices under investor pressure.
Markets brace for a grind
Michael Burry is calling for a long selloff that could outlast the dot-com bust, citing frothy AI plays, the rise of passive funds, and accounting that hides the true cost of AI hardware. On the labour front, US companies announced 71,321 layoffs in November, bringing the 2025 total to 1.17 million - the highest since 2020 - even as unemployment hovers near 4.1% and holiday spending looks resilient.
Attention, addiction, and the 10-minute test
Aadit Sheth flags a large meta-analysis linking heavy short-form video use with weaker attention and impulse control. His simple bar is telling - if you can sit with a hard problem for 10 minutes without reaching for your phone, you are already doing what many cannot.
What counts as open source
DHH defends calling 37signals’ Fizzy “open source” under an MIT-derived O’Saasy Licence that blocks third-party SaaS clones while allowing self-hosting and contributions. Critics argue that restrictions make it source-available, not open under OSI rules - a fight that echoes recent licence pivots by big projects.
Tech for money, or tech as maths
Balaji draws a line between chasing “tech” for profit and engaging with the maths that underpins real progress. Replies connect it to crypto versus cryptography and other fields where branding outruns substance.
Politics of fraud and oversight
Leading Report notes Congress probing Minnesota Governor Tim Walz over the $240 million Feeding Our Future scandal, part of COVID-era nutrition funds abuse. Expect this to play into 2026 state politics as court cases work through the system.
Fast trains, fuller context
@levelsio’s China rail thread sparks a speed comparison with Japan’s Shinkansen, plus a cultural note - business class lounges can look empty because most riders favour cheaper, perfectly fine second-class seats on 350km/h lines.
What a “reasoning model” is - and isn’t
Teknium reminds people that Mistral Large 3 is a non-reasoning MoE model tuned for speed and cost - so judging it on chain-of-thought benchmarks misses the point. The debate underlines that not every app needs deep reasoning, some need low latency and predictable output.
Why it matters
We are moving into a practical phase for AI - faster voice on modest hardware and reasoning modes inside mainstream apps. That widens access, but also raises questions about safe use, watermarking, and how to measure “reasoning” in the wild.
Workplaces are renegotiating their relationship with AI. People want gains without losing identity-defining tasks, and stigma still dogs creative fields. Vendors that listen - with real pilots and mixed methods - will earn trust faster.
Apple’s handover shows incumbents are not immune to missed bets. Leadership changes are a signal to ship, not a fix in themselves.
Markets are testing the gap between story and cash flows. If hardware costs and passive flows have masked fragility, then earnings seasons ahead will be the referee. Layoff announcements tell you where leaders think the economy is heading, not where it is today.
Attention is a competitive edge. The 10-minute rule is simple, cheap, and a good early-warning check on your own habits.
Licences and labels matter. Calling source-available “open” compresses a real debate about sustainability and freedom. Teams should read the fine print before building on any stack.
Finally, the culture wars inside tech - maths versus memes, reasoning versus speed - will shape which tools win. Clarity about goals beats brand chasing every time.
