- AI Fire
- Posts
- 🔥 GPT-5.2 Is Coming
In partnership with
Anthropic just ordered almost a million Google TPUs. OpenAI is testing a secret model called Garlic. And Chinese labs claim a 14B coder that can hang with o3. Are we watching GPUs, big models, and even Nvidia’s moat get rewritten in real time…!?
What are on FIRE 🔥
IN PARTNERSHIP WITH SYNTHFLOW
The New Framework for Enterprise Voice AI
Enterprise teams are automating more calls than ever — but without a consistent framework, deployments become unpredictable, costly, and slow to scale.
The BELL Framework introduces a structured way to design, test, launch, and improve Voice AI agents with reliability.
Get the guide enterprises are now using to de-risk voice automation and accelerate deployment.
AI INSIGHTS
Google is finally selling TPUs at scale. And Anthropic is buying big enough to shake Nvidia.
1. Anthropic’s massive order
About 1M TPUv7 chips.
400k bought as full racks
600k rented through GCP
Built by Broadcom, deployed with Fluidstack + TeraWulf
This turns Google into a true merchant hardware vendor, not just “TPUs for Google only”.
2. TPUs beat GPUs on cost
Specs aside, TPUs win where it matters.
~30–40% lower TCO than GB200/GB300
Anthropic can get ~50% lower cost per effective PFLOP
Even OpenAI saved ~30% by just threatening to switch
3. The software shift
Google is fixing its main weakness.
Native PyTorch on TPU
vLLM / SGLang support
New kernel tools like Pallas
Only big gap left. Google still hasn’t open-sourced XLA:TPU.
Why it matters: Nvidia is no longer the only serious option for frontier labs. Anthropic, Meta, xAI, OpenAI now have real pricing leverage. If Google opens the TPU stack fully, the real moat shifts from “who has GPUs” to “who owns full systems, compilers, and power.”
PRESENTED BY LEVANTA
The Future of Shopping? AI + Actual Humans.
AI has changed how consumers shop, but people still drive decisions. Levanta’s research shows affiliate and creator content continues to influence conversions, plus it now shapes the product recommendations AI delivers. Affiliate marketing isn’t being replaced by AI, it’s being amplified.
AI SOURCES FROM AI FIRE
1. AI trading playbook for 2025. 8 ways to turn ChatGPT and Claude into your personal analyst, coder and risk coach so you stop guessing and start trading with a system
2. Build a full business with FREE Google AI. Copy the exact NotebookLM → Firebase → Pomelli → Veo 3.1 workflow to launch a research-backed website, brand and video ads in 1 day without paying for tools
FINAL 12-HOUR EXTENSION REMINDER
It’s more than this, but we can’t put all on a screen
Like we said just 2 days ago, we woke up with full of messages from all of you who missed the Black Friday deal.
After seeing the volume of requests & getting an emergency call, we agreed to open the offer for a little longer.
We are extending the Black Friday discount until today, Dec 7th, 23:59 EST. And after that, it is gone for good!
If you wanted the 30% off for our biggest Livfetime update at no extra cost, this is the last window. All coming soon. Getting in now puts you in the best position for 2026.
TODAY IN AI
AI HIGHLIGHTS
🔥 OpenAI just pulled a “code red” and rushed out GPT-5.2, its answer to Google’s Gemini 3 beating the leaderboards. Early tests say it closes the gap and might even edge ahead on reasoning. And it could ship as early as next week.
📉 Investors are starting to ask tough questions about OpenAI’s spending, losses, and slowing ChatGPT growth. Meanwhile, Alphabet is soaring as Gemini momentum lifts every partner stock around it. The AI trade is shifting fast.
🛡️ A nonprofit just published the first AI Safety Index and even top labs only earned a C+. OpenAI and Anthropic scored highest. Google trailed. Meta, xAI, and Alibaba fell to the bottom. The big headline. Safety is still nowhere near where it should be.
🧄 OpenAI’s next big swing is Garlic, a new model that fixes pretraining issues and already beats Gemini 3 and Opus 4.5 on internal coding and reasoning tests. It might land as GPT-5.2 or even 5.5 early next year. Internal teams say this one is a real jump.
💸 Chinese models keep winning on cost. DeepSeek V3.2 now matches frontier models on intelligence while being up to 22× cheaper to run. Tool-optimized, reasoning-optimized, open-weight. And it is putting serious price pressure on Google, OpenAI, and Anthropic.
💰 Hot AI Acquisition: Anthropic just acquired Bun after Claude Code hit a $1B run-rate in only 6 months. Bun’s ultra-fast JavaScript runtime, trusted by Midjourney and Lovable, brings 7M+ monthly downloads and 82K GitHub stars. This deal strengthens Anthropic’s lead in enterprise AI and developer tools.
NEW EMPOWERED AI TOOLS
🎥 TruGen AI builds human-like video agents that see, hear & act in real time for lifelike interactions.
🗺️ Eureka turns any topic into a visual knowledge map, transforming PDFs or ideas into interactive graphs.
🥗 Bean is the AI meal planner for families, creating personalized menus, grocery lists & recipe collections.
🌐 Browser Use makes websites fully accessible to AI agents, extracting interactive elements for automation.
AI BREAKTHROUGH
50 researchers from Bytedance, Alibaba, Tencent, and top universities just dropped a 300-page deep dive into how coding models actually learn. It’s one of the clearest looks we’ve seen into why small open models suddenly feel smarter. And why giant frontier models still fail at basic software tasks.
Small > big with RLVR. A 14B RLVR model matches OpenAI’s o3.
Python is hard. Static languages boost each other. Python breaks the pattern.
Languages scale differently. C# and Java need less data. Python and JS need more.
MoE = power + instability. Huge capacity, fragile training.
Models learn insecure code. Public repos teach unsafe patterns.
Reasoning is structural. CoT teaches the template, not the facts.
“Lost in the middle” stays. Long context doesn’t fix it.
Debate helps if tuned well. Bad setups cause worse hallucinations.
Alignment tax is real. Safety tuning reduces coding skill.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply