- NATURAL 20
- Posts
- OpenAI Researchers Quit Over Hidden Risks
OpenAI Researchers Quit Over Hidden Risks
PLUS: China Builds Secret Chip Lab, Amazon AI Chief Steps Down and more.
Wes Roth
December 18, 2025
In partnership with
Your competitors are already automating. Here's the data.
Retail and ecommerce teams using AI for customer service are resolving 40-60% more tickets without more staff, cutting cost-per-ticket by 30%+, and handling seasonal spikes 3x faster.
But here's what separates winners from everyone else: they started with the data, not the hype.
Gladly handles the predictable volume, FAQs, routing, returns, order status, while your team focuses on customers who need a human touch. The result? Better experiences. Lower costs. Real competitive advantage. Ready to see what's possible for your business?
Hey friends,
Welcome back to the newsletter! I hope you’ve had plenty of coffee (or tea) this morning because the AI world clearly hasn't been sleeping.
We’ve got Google dropping a new generation of speed, OpenAI seemingly looking for a new massive allowance, and Mistral getting a little artsy.
Let’s dive into the good stuff.
Today:
OpenAI Researchers Quit Over Hidden Risks
Google Releases Gemini 3 Flash
OpenAI Talks $10B Amazon Deal
Mistral Launches Creative Model
China Builds Secret Chip Lab
Amazon AI Chief Steps Down
OpenAI's "SECRET RESEARCH" Revealed
Multiple researchers have left OpenAI, accusing the company of hiding research on AI’s economic risks. One of them, economist Tom Cunningham, said the team had shifted from real research to promoting company interests. The resignations came just before OpenAI released GPT-5.2, which outperforms experienced professionals on complex tasks—suggesting real threats to white-collar jobs, especially early-career roles.
Meanwhile, rival Anthropic has been open about AI’s impact, warning of a “white-collar bloodbath.” Critics say OpenAI is becoming more secretive ahead of a possible IPO, raising concerns about the public's ability to assess AI’s true societal costs.
The big headline: Amazon is reportedly in talks to invest about $10B (or more) into OpenAI, and part of the discussion includes OpenAI using Amazon’s Trainium AI chips. The reporting also frames this as potentially valuing OpenAI north of $500B, though talks are described as fluid.
Why this feels like a real shift (if it lands): it’s not just money — it’s supply chain + leverage. If OpenAI becomes meaningfully “multi-cloud / multi-chip,” that changes negotiating power with everyone (clouds, chip vendors, and partners). It also puts Trainium in the conversation as more than “AWS internal tooling.”
My practical takeaway for builders: keep an eye on whether this turns into real, available capacity and better pricing (especially for inference) or stays mostly strategic. Either way, it’s another signal that compute access is the product.
Google’s calling Gemini 3 Flash “frontier intelligence built for speed,” and they’re rolling it out through the Gemini API (AI Studio), Antigravity, Gemini CLI, Android Studio, and for enterprise via Vertex AI.
A few details that jumped out:
Benchmarks (as reported by Google): GPQA Diamond 90.4%, and Humanity’s Last Exam 33.7% (without tools).
Speed/efficiency: Google says it’s ~3x faster than 2.5 Pro (citing Artificial Analysis benchmarking) while outperforming 2.5 Pro.
Pricing (Gemini API / Vertex AI): $0.50 / 1M input tokens and $3 / 1M output tokens (with audio input noted separately).
Cost levers: built-in context caching (Google says up to 90% reductions in repeated-token scenarios) and a Batch API (Google says 50% cost savings + higher async rate limits).
Multimodal “do stuff” feature: it now offers code execution to zoom, count, and edit visual inputs.
And I liked that they gave concrete early use cases: near real-time deepfake analysis claims (Resemble AI says 4x faster multimodal analysis vs 2.5 Pro), plus legal doc analysis (Harvey), plus game creation workflows.
If you want a lighter, story-friendly model to toss into your workflow: Mistral Small Creative is now listed on OpenRouter with 32,768 context, priced at $0.10/M input tokens and $0.30/M output tokens.
OpenRouter describes it as an experimental small model aimed at creative writing, narrative, roleplay, character dialogue, and general instruction following.
If you’re doing scripting, dialogue passes, or “generate 30 variations and pick 3,” this kind of pricing + context is exactly where I like to start before I escalate to heavier frontier calls.
If you tell me what you’re building this week (newsletter, scripts, agents, internal tool, etc.), I can suggest which of these to try first and a couple prompts that tend to separate “looks good” from “actually usable.”
🧠RESEARCH
This paper introduces a new system for creating long, high-quality avatar videos that don't get blurry or glitchy over time. It builds videos in stages to ensure details remain sharp and uses specialized software "directors" to keep characters looking consistent. It solves the common problem of AI videos losing quality as they get longer.
Current AI can make realistic images, but does it understand what it's creating? The MMGR project tests top models on physics, logic, and 3D space. The results show that while these AIs create beautiful visuals, they often struggle with basic common sense, like how objects should move or interact in the real world.
Researchers created a difficult test using detailed ASMR videos to see if AI can trick humans and detection software. They found that the latest AI video generators are convincing enough to fool most automated tools. However, humans are still better at spotting fakes, especially when sound is included to reveal subtle mistakes.
🛠️TOP TOOLS
Each listing includes a hands-on tutorial so you can get started right away, whether you’re a beginner or a pro.
AstroGPT AI – Verdic Astrology - AI astrologer that blends classical Vedic astrology with large‑language‑model reasoning.
Audiosonic – AI Voice Generator - AI voice generator that converts written text into natural‑sounding speech
Audo Studio – AI Audio Cleaning Tool - browser‑based tool that cleans up speech in seconds—removing background noise and leveling your audio automatically.
📲SOCIAL MEDIA
🗞️MORE NEWS
China’s Secret Chip Lab China has secretly built a massive laboratory to recreate advanced chip-making machines, aiming to break free from U.S. trade restrictions. They hired former engineers from ASML, the leading Dutch equipment maker, to copy complex tools that use light to print tiny circuits on computer chips. While they have built a prototype machine, it is not yet capable of mass-producing the powerful processors needed for modern AI.
Amazon AI Leadership Shakeup Amazon’s head of artificial general intelligence (AGI) is leaving, triggering a major reorganization of how the tech giant builds its smartest AI. He will be replaced by a veteran executive who will now oversee AI software, custom computer chips, and futuristic quantum computing all in one unified group. This move signals that Amazon wants to tightly link its brainy AI models with the specialized hardware that runs them.
Google’s New AI Coding Tool Google launched a new toolkit that lets programmers build autonomous AI "agents" using TypeScript, a very popular coding language. Instead of just asking the AI to do things with text prompts, developers can now write structured code to control exactly how these digital helpers behave. This "code-first" method makes it much easier to test the AI and ensure it works correctly when building complex software.
Roblox Tests AI Game Builders Roblox created a new testing system called OpenGameEval to measure how well AI can actually help people build video games. Unlike simple coding tests, this benchmark challenges AI to complete multi-step tasks like fixing game rules or changing 3D objects in a virtual world. The goal is to identify which AI tools are truly useful for game creators rather than just good at writing short snippets of text.
A Mathematician’s View on AI Famous mathematician Terence Tao argues that today's AI isn't truly intelligent but rather possesses a "jagged" kind of cleverness. He suggests treating these tools as random idea generators that can offer brilliant solutions, provided humans carefully check the work for errors. This perspective shifts the focus from expecting human-like understanding to using AI as a powerful, albeit imperfect, brainstorming partner.
The AI Influencer Fake-Out A hacker revealed that a startup funded by top investors is running a "phone farm" of over 1,000 physical smartphones to fake popularity on TikTok. The company used these phones to manage fake, AI-generated "influencers" that posted ads without telling viewers they were advertisements. This breach exposed how some businesses are secretly flooding social media with artificial content to trick algorithms and users.
What'd you think of today's edition? |
Reply