- AI with Kyle
- Posts
- AI with Kyle Daily Update 103
AI with Kyle Daily Update 103
Today in AI: ChatGPT Garlic model + Mistral 3
Kyle Balmer
December 04, 2025
The skinny on what's happening in AI - straight from the previous live session:
🧄 OpenAI Code Red: New Model "Garlic" Coming Next Week
Altman pausing ads, agents, everything. All hands on deck to beat Gemini 3. Garlic to ward off the vampire.
Kyle's take: More details on panic mode from Monday.
They're calling new model "Garlic" - perhaps to ward off Gemini the vampire??
Garlic is supposedly coming next week, supposedly beats Gemini 3 in internal tests. It is apparently a fine-tuned model. We’ll find out soon.
What’s more important here is that everything else PAUSED - ads delayed, AI agents stopped, health assistant frozen. All compute and people redirected to core ChatGPT.
Priority areas: better personalisation (after GPT-5 backlash from 4.0 lovers), faster responses, fewer "unnecessary refusals" (remember chocolate melting instructions triggering suicide hotline?).
Image generation has also been named a priority but good luck - Google has Google Images, decades of labeled data. OpenAI scraping and buying data while Google already “owns” it.
Two years ago Google was panicking about ChatGPT. Tables completely turned. Competition's finally here and OpenAI wasn't ready. They thought they had years of runway. Oops.
My take: great! Competition is good. We need Sputnik moments when complacency gets its comeuppance.
🌍 Mistral 3: Europe's AI That's Slower, Dumber, and More Expensive Than Everyone
Mistral releases a model - and it promptly gets teased as an. EU model "within striking distance of GPT-4"… But their 3B browser model is genuinely cool!
Kyle's take: Poor Mistral. Released day after DeepSeek. Analysts brutally say it's "finally catching up to DeepSeek from 2023."
On benchmarks: dumb, slow, expensive. They hit NONE of the three things you want (smart/fast/cheap). Really hard to think of use cases, except within EU companies that require an EU AI. It’s opensource but nowhere near as good as DeepSeek 3.2.
Europe's out of the race - not enough money, compute, or electricity.
BUT Mistral's tiny 3B model runs IN YOUR BROWSER. I had it watching me live, describing my expressions in real-time. Three gigabytes in Chrome cache, totally private. This may be the the play for everyone non-Us, non-China - edge AI, efficiency over brute force. Europe can't compete on size but could at least get into the game on devices and efficiency.
HIGHLY recommend playing with the browser demo. It’s extremely impressive.
Source: Mistral announcement, browser demo
đź’° Anthropic IPO 2026: First big AI Company Going Public?
Beating OpenAI to market? Hired IPO lawyers, eyeing acquisitions. Getting out before bubble pops?
Kyle's take: Anthropic hired Wilson Sonsini for IPO prep - could go public 2026, beating OpenAI's 2027 target. First major AI company to IPO.
They also just bought Bun and potentially looking for more acquisitions?
Here's my theory: they're rushing to IPO before bubble pops. Get liquidity for investors, cash out employees, then weather whatever storm's coming. Smart move. And get it done before OpenAI’s IPO.
That said Anthropic are also getting the work done. While OpenAI's having code reds and panic attacks, Anthropic just grinds away releasing best models with zero hype. Dropped Opus 4.5 (best model on Earth) like ordering coffee. No conference, no fanfare, just tweet saying "it's ready, here you go."
Source: Reuters
🟡 The Mystery of ChatGPT's "Piss Filter" Finally Solved?
Why does every ChatGPT image look yellow? ChatGPT trained on its own images after taskers fed synthetic data back into the model.
Kyle's take: I found this discussion of why ChatGPT images have that weird yellow hue - the "piss filter" as people call it (mean but accurate).
Here's what supposedly happened: OpenAI went viral, needed labeled data fast. Hired taskers in Philippines/Argentina to find images matching prompts. But ChatGPT images were already everywhere online. Taskers, trying to get paid quickly, grabbed ChatGPT images from Google Images and fed them BACK into training. ChatGPT essentially ate its own tail. The whole image model got poisoned off the back of this, making everything that weird yellow.
But why yellow? It could have been any colour really. Once the synthetic data starts poisoning itself the bias could have gone in any direction.
My guess? Probably from Studio Ghibli style images - that warm, nostalgic, autumnal feeling. Once enough yellow-tinted synthetic data 9 (ie. the HUGE influx of Ghiblified images) entered training, the model locked onto it. Now EVERYTHING has that hue.
Unless they retrain from scratch (they won't), we're maybe stuck with piss-colored ChatGPT images forever. Meanwhile Google has decades of properly labeled Google Images data. No wonder Nano Banana looks better….
🎮 Google's Amazon Strategy: Bleed Everyone Dry, Then Monetise
Keep Gemini free until OpenAI dies. It worked for Amazon with retail.
Kyle's take: I’m seeing lots of people celebrating Google becoming a threat to OpenAI. I’m also happy we have competition.
However: Google could pull an Amazon. Remember Amazon lost money for ~10 YEARS, choking out bookstores, then retail, then everything? They just undercut, ran at a loss and outlived the retail industry using investment as a float. Once they’d killed traditional retail (the competition) they shifted gears and began to make a profit.
Google has the cash to do same with AI. Keep Gemini free, no ads, pure AI experience. OpenAI bleeding money, needs funding rounds to survive. OpenAI introduces advertisements to staunch the blood loss - annoys customers who switch to Gemini.
Google just... waits. No rush. No hurry. Like python squeezing prey. They own the entire stack - chips, data centres, cloud, apps, users. And oodles of cash.
OpenAI owns a chatbot and hype machine. Once OpenAI collapses under its weight, Google adds ads, raises prices, owns everything.
OR China does this with open source - DeepSeek already 20x cheaper. Either way, OpenAI's business model is shakey. Too big to fail? We've been there before.
(Not investment analysis/recommendation btw- consider this my fanfic!)
Member Questions:
Kyle's response: Main reason is Gemini 3. But it's also unfair - ChatGPT's been default so long we've found all its problems. New models like Gemini feel better because we haven't hit the rough edges yet. It's novel, exciting. The grass is always greener!
But there are legitimate concerns with OpenAI too - privacy especially. Google's handled our data for decade+, mostly responsibly (yes ads, but not wholesale selling). OpenAI? No track record, arguably negative track record. Better the devil you know.
Kyle's response: Probably next week. Don't know if 5.2 but something called "Garlic" - to ward off Gemini the vampire. OpenAI reps on social media hyping "wait till you see what's coming." They need something big for 12 Days of Shipmas or they're toast. Garlic on toast? Yum!
"Love Claude's focus, OpenAI should do same" Comment about approach
Kyle's response: Claude's hyper-focused. When they released Opus 4.5, just tweeted "here it is." If OpenAI released it, huge conference, "best thing ever!" Claude under-markets because product speaks for itself. Very good at coding and work tasks. No images, no video, voice mode's rubbish. But devastatingly effective at core function.
Want the full unfiltered discussion? Join me tomorrow for the daily AI news live stream where we dig into the stories and you can ask questions directly.