• AI Fire
  • Posts
  • 🧄 GPT-5.3 “Garlic” Leaked Overview

🧄 GPT-5.3 “Garlic” Leaked Overview

Seeing in 4D with Google's D4RT!!!

In partnership with

ai-fire-banner

A new benchmark shows top AI agents still can’t handle real office work, so… don’t worry just yet. Google just revealed a model that sees the world in 4D like a human.

IN PARTNERSHIP WITH BELAY

AI promises speed and efficiency, but it’s leaving many leaders feeling more overwhelmed than ever. The real problem isn’t technology. It’s the pressure to do more with less without losing what makes your leadership effective.

BELAY created the free resource 5 Traits AI Can’t Replace & Why They Matter More Than Ever to help leaders pinpoint where AI can help and where human judgment is still essential.

At BELAY, we help leaders accomplish more by matching them with top-tier, U.S.-based Executive Assistants who bring the discernment, foresight, and relational intelligence that AI can’t replicate.

That way, you can focus on vision. Not systems.

AI INSIGHTS

are-ai-agents-office-ready-not-yet

You’ve heard it all year: “AI is ready for real jobs.” But when it comes to actually doing professional work, they’re not ready yet.

None of them passed APEX-Agents, a benchmark that drops today’s best models into real white-collar workflows.

They’re the kind of cross-domain tasks lawyers or analysts deal with across Slack, Google Drive, and dense policy docs. So, how’d the models do?

  • Gemini 3 Flash: 24% accuracy

  • GPT‑5.2: 23%

  • Opus 4.5 / GPT‑5 / Gemini 3 Pro: ~18%

  • Everyone else: lower

In most cases, the models either got the answer wrong or just gave up. Even top-tier models couldn’t piece together multiple sources of information or reason across them. And that’s the core of knowledge work.

For all the “AI will replace knowledge workers” talk, this is the first real test that simulates the job, and even the best models still perform like interns on their first week.

🎁 Today's Trivia - Vote, Learn & Win!

Get a 3-month membership at AI Fire Academy (700+ AI Workflows, AI Tutorials, AI Case Studies to automate your work, 2x your free time) just by answering the poll.

What Caused Most Models to Fail?

Login or Subscribe to participate in polls.

PRESENTED BY ATTIO

Introducing the first AI-native CRM

Connect your email, and you’ll instantly get a CRM with enriched customer insights and a platform that grows with your business.

With AI at the core, Attio lets you:

  • Prospect and route leads with research agents

  • Get real-time insights during customer calls

  • Build powerful automations for your complex workflows

Join industry leaders like Granola, Taskrabbit, Flatfile and more.

AI SOURCES FROM AI FIRE

1. GPT-5.3 “Garlic”: The first leaked comprehensive preview. What to expect? Here’s why it wins on cost, speed, reasoning. Save this to see if it’s true later

2. Step-by-step guide to create consistent & full slides with No PowerPoint. We create beautiful decks using simple scripts. No more dragging boxes or fonts

3. NotebookLM became an all-in-one researcher. Why everyone is sleeping on this Gem? I tested it again and found 7 real workflows that turn raw research into slides, audio and tables in minutes

4. Stop writing proposals: Turn sales calls into decks in 3 minutes with AI (Free Template). Here’s the exact n8n flow to build a multi-agent AI system that listens to your sales calls and delivers a 90% finished Gamma proposal to your inbox (copy & paste flow if you don’t want to read)

TODAY IN AI

AI HIGHLIGHTS

🖥️ If you opened Claude Cowork and didn’t know what to do next, just watch this simple 13-min exercise to start moving & turn Cowork into your thinking partner.

→ After that, I highly recommend you to try adding these 10 essential workflows (+ prompts) to automate your life’s most boring tasks with Claude Cowork.

🧪 Leaked: OpenAI’s internally testing a new tool code-named “Salute”. It lets you upload files, assign tasks, and track progress. You can see some first previews here.

🎵 ElevenLabs just dropped a 13-track AI album featuring Liza Minnelli, Art Garfunkel & others. Full royalties, no lawsuits, and labels are actually in. Maybe give it a listen?

🎧 “Make me a playlist that feels like rainy Tokyo nights.” Spotify’s new Prompted Playlists truly understands that now. Just describe a vibe, it’ll build a full list for you.

📉 Anthropic’s margins just dropped from 50% to 40%. As Google + Amazon jacked up server costs by 23%, cutting inference profits (according to The Information)

💰 Big AI Deal: Saudi Infra Fund & Humain agreed on a financing deal of $1.2B to develop AI data centers, showing a major investment in advancing AI infrastructure

NEW EMPOWERED AI TOOLS

  1. 📊 ChartGen AI turns raw data (all different data sources, from Facebook to TikTok) into professional charts with insights in seconds

  2. 📍 LocateStore creates a map of all your stores using Google Sheets. Add an address in a Sheet, and get an interactive map with search and filters

  3. 🎥 Demonstrate records any browser task once, gets production-ready code & deploys automation code as a serverless function

  4. 📅 Callum (AI calendar assistant) gets multi-person meetings, shared availability, and real-world scheduling constraints

  5. CyberCut AI auto-slices long footage into social-ready clips, generates marketing videos, adds high-precision subtitles

AI BREAKTHROUGH

googles-4d-aware-vision-model-can-see-like-us

You’ve heard of AI that chats. AI that codes. Even AI that makes music. But what about AI that sees? Not just seeing a picture. It tracks every pixel across time, reconstructs full 3D scenes from 2D video,… That’s what D4RT does.

It stands for Dynamic 4D Reconstruction and Tracking, and it’s a unified AI model that does something that used to take a clunky stack of tools:

  • Rebuild 3D geometry from a flat video

  • Track objects in motion, even if they leave the frame

  • Estimate camera position from any angle

  • Handle occlusion, motion blur, and real-time chaos

Instead of training a dozen separate models, D4RT turns all of that into a single, efficient query-based system.

The model builds a compact internal “world model” from the video, then answers that query. That’s why D4RT is 18× to 300× faster than anything before it. It gets parsed a one-minute video in only 5 seconds.

(Previous models took 10 minutes to do the same.) And because it only computes what it’s asked for, it’s efficient enough to run in real time.

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

Login or Subscribe to participate in polls.

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

or to participate.