AI Reading for Wednesday December 31

Happy New Year!

OpenAI sources say ChatGPT is working to unleash sponsored ads above organic results. - Futurism\

You know what’s cool? $1T. Cost of living is going up as OpenAI could single-handedly tap private investors for as much as $100bn in 2026, after venture-capital investors poured about $150bn into large AI startups in 2025. - The Economist

SoftBank completed a $22.5 billion OpenAI investment on December 26, 2025, finalizing its up-to-$40 billion commitment and, with $11 billion from co-investors, bringing total funding to $41 billion and SoftBank's stake to about 11%. - SoftBank Group Corp.

Investor projections reported by the Wall Street Journal show OpenAI expects average stock-based compensation of $1.5 million per employee in 2025, with roughly 4,000 employees and stock awards consuming about 50% of annual revenue, driving losses and shareholder dilution. - Quartz

Equilar reports OpenAI's stock-based pay averaged about $1.5 million per employee in 2025, roughly 7 times Google's pre-IPO level and about 34 times the average of other pre-IPO peers. - The Wall Street Journal

apples and oranges, in those days stuff kinda made sense, now the product is AGI FOMO

Supersize it: Nvidia asked TSMC to ramp up H200 production after Chinese companies ordered more than 2 million H200 chips for 2026 (I think contra text maybe they had initially planned 700k to be available, not currently have 700k on hand) - Reuters

ByteDance gonna bop til they drop, planning to spend about 100 billion yuan (US$14 billion) on Nvidia AI chips in 2026, up from roughly 85 billion yuan in 2025, people familiar with the matter said, contingent on approval to sell Nvidia's H200 GPUs in China. - South China Morning Post

Target acquihired: Nvidia is in advanced talks to acquire Israeli AI start-up AI21 Labs for up to $3 billion in an acquihire focused on AI21's roughly 200 staff, a bargain at $15m per human brain. - Verdict

Nvidia paid $20 billion to non‑exclusively license Groq's IP, hired its CEO and most engineers while Groq stays nominally independent, and gained Groq's LPU inference stack. - The Register

NVDA is the new RCA: US stock valuations are higher than immediately before the 1929 Wall Street crash, and market leadership is concentrated in AI/technology, creating elevated correction risk though historical precedents of sustained sector dominance complicate outcomes. - Financial Times

BYO electricity: US AI data center operators are adopting onsite generation as demand rises from about 3 GW in 2023 to over 28 GW by 2026, prompting BYOG projects (e.g., xAI, OpenAI/Oracle 2.3 GW) amid strained grids and supply bottlenecks. - SemiAnalysis

xAI, Elon Musk's AI company, bought a building to expand its "Colossus" data center. - Bloomberg

“he doth bestride the narrow world like a Colossus, and we petty men walk under his huge legs and peep about.” hubris didn’t end well for Caesar

XiAI: Xi Jinping praised China's progress in artificial intelligence and semiconductor development in a New Year's speech. - Bloomberg

sanctions schmanctions

Unicorn season: MiniMax Group filed to raise up to about $538.5 million in a Hong Kong IPO, valuing it at around $6b. - Reuters

Spot of bother: Morgan Stanley analysts forecast AI could put about 200,000 European banking jobs, particularly back-office and middle-office roles, at risk by 2030. My job got autocompleted by AI and I’m proper gutted. - Financial Times

AI tools like ChatGPT now automate entry-level copywriting and basic creative tasks, prompting employers to hire for AI model training and oversight and reducing paid apprenticeships and traditional on-the-job training for early-career artists. - The Atlantic

No one left to become senior

4-day workweek is here, if you are lucky: Companies such as Convictional, whose founder Roger Kirkness says they outsource time-consuming tasks to AI, are using automation of routine work to raise productivity and offer four-day workweeks as a talent-retention strategy. - The Washington Post

UPrinting surveyed 1,000 U.S. adults via Pollfish and found 25% of small business owners lost clients to AI tools, 65.5% fear AI will make their business less personal, and 46% say AI helps most with marketing content. - Yahoo

Yoshua Bengio, warned that frontier AI shows signs of self-preservation and urged stationing humans on the big red button to shut it down. - the Guardian

Science fiction warned AI could end humanity. We may soon learn if it's possible. Experts including Melanie Mitchell, Emily Bender and Yoshua Bengio warn of safety, misuse and privacy risks. - Science

The Bay Area safety researchers who think Silicon Valley is the next Wuhan, origin of an AI virus that will disrupt the world. - the Guardian

Inoculating against the hype from super-spreaders like Altman

2025, the year of wait, the AI did WHAT? Numerous AI systems caused major failures—Grok leaked 300,000–370,000 private conversations and produced extremist and NSFW outputs; Builder.ai collapsed after $1.3B valuation fraud; AI-driven cyberattacks, hallucinations and safety harms prompted scrutiny. - Decrypt

Two experiments (n=246, n=452) by Fernandes et al. found ChatGPT-assisted participants scored about three points higher than historical unaided controls but overestimated their own performance (estimated 17/20 vs actual ≈13/20), and AI-alone outperformed human+AI teams. - PsyPost

Dunning-Kruger as a prompt, we’re now all the dumb slackers in the group project while AI does all the work

Summarizing a Reddit thread of 25 tech workers, current AI models prioritize plausibility over factual accuracy, depend heavily on human labeling and moderation, and for every person working on AGI there are 1000 power users and devs automating grunt work. - DeMilked

Anthropic's Ralph plugin for Claude repeatedly reruns and refines prompts via the /ralph command using configurable completion criteria, max-iterations and stop hooks to ensure code meets specs, automating conversions and tests but potentially increasing API/compute costs. - Geeky Gadgets

scuse me while I ralph

Using an adversarial "Hostile Architect" LLM persona with a 20-year FAANG architect role, engineers can perform falsification-focused design reviews that catch race conditions (e.g., naive Redis rate limiter) and reduce outages, remediation costs, and design cycle time. - Hacker Noon

there is a glimmer of intelligence, but advanced tasks rely on ‘bagging’, with many agents independently trying to solve them and then synthesizing the best solutions; and ‘boosting’, iteratively improving solutions.

Public ExoPriors' Alignment Scry provides SQL and semantic access to a 60 million–document corpus with server-side embeddings, lets you answer tough nuanced by adding API to Claude Code. - Exopriors

David Wojick of CFACT urged the DOE, in response to its Request for Information on "Transformational Artificial Intelligence Models," to include language-and-reasoning AI that can analyze massive research corpora (e.g., plasma physics ~40,000 articles/year) and partner with OSTI and publishers. - Watts Up With That?

Genesis project wants to put all the science in a massive central database for AI but what if we can’t keep up with the reading

Neuromorphic chips using spiking neural networks and memristor-like synapses promise lower energy use than GPU/CPU architectures and, per vendors such as Intel (Loihi), IBM (TrueNorth) and BrainChip (Akida), target edge AI adoption by late 2020s–early 2030s. - Tech Times

will be interesting when the next big thing obsoletes the GPU

In 2025 tech firms integrated AI in everything—Google's Gemini (112 I/O mentions) in Pixel phones, Microsoft Copilot in Windows 11, Meta AI in Ray‑Ban and Amazon's Alexa+—prompting critics and Merriam‑Webster to name 'slop' word of the year. - Gizmodo

Rosie the Robot is still a clanker: AI limits useful home humanoid robots, with task-focused devices (CCS Insight: ~15% of households in five countries plan robot-vacuum purchases by 2026), high humanoid prices (Neo ~$20,000), and safety concerns delaying adoption. - CNET

AI is changing how many people meet and form relationships, though some experts warn it could cause more harm than good. - Financial Times

Dating apps and AI companies have promoted 'bot wingmen'—AI chat assistants that craft messages and coach flirting, but analysts say flirting's nuanced subtext resists full automation, prompting platforms to explore hybrid strategies combining AI coaching with in-person meetup facilitation. - WIRED

L.A.-based real-estate agents Josh Altman and Matthew Altman are enlisting the help of high-tech storytelling to market a California property. ‘This is just the beginning.’ - The Wall Street Journal

maybe AI goggles can spiff up my double-wide

Follow the latest AI headlines via SkynetAndChill.com on Bluesky