Another week in AI means more breakthroughs, new models, incredible research, and massive leaps in hardware. I’ve scoured a lot of dark and obscure place of the interwebs to bring you this content as usual.
TL;DR
This week in AI was like trying to drink from a firehose. There’s GPT-4.5 teasing us, Meta throwing llamas everywhere, Nvidia releasing chips that could probably launch a rocket, and AI in healthcare diagnosing stuff better than your doctor. Oh, and robots learning faster than I can make coffee. But that is not a real feat, cause my OutIn Nano just broke down after just 6 months.
- OpenAI teased GPT-4.5—this one’s faster, smoother, and ready to have real conversations.
- Meta’s LLaMA-3.0 hit the stage, claiming to be the open-source GPT-4 rival.
- Adobe Firefly is now in Photoshop and Illustrator, meaning you’ll never have to learn real design skills again.
- Nvidia released the H100 Superchip, designed to process AI workloads at lightspeed (and maybe build your dream house).
- DeepMind unveiled Gemini, a reinforcement learning model for robots. Get ready for smarter, faster robots—just like sci-fi, but without the lasers (yet).
- IBM WatsonX is diagnosing medical images faster than humans, and hospitals are loving it.
- AI startups are raking in billions. Anthropic got $1 billion, and Elon Musk’s xAI built Colossus, a supercomputer with 100,000 Nvidia GPUs. It’s like Skynet, but with better branding.
But let’s face it, a lot of this is still in early stages. It is cool tech though, but it aint yet changing our lives.
But still….
Let’s get into the details.
Before we start!
You know the drill:
- Like this post, because Google rewards engagement like it’s handing out candy. 📢
- Connect with me on LinkedIn and let’s be buddies. 🔗
- Subscribe to my TechTonic Shifts newsletter to get these updates without having to scroll through your feed. Seriously, your inbox will thank you. 💉
Main Spotlight
OpenAI’s GPT-4.5: Real-Time Reasoning for Dynamic Conversations
OpenAI just dropped GPT-4o1, but let’s be real, we’re all waiting for GPT-4.5.
Why you ask…
Because 4.5 is going to be the chatty, quick thinker we need—learning as it goes and making conversations feel more natural. While details are still hush-hush, it’s expected to be a game-changer for chatbots, customer service, and any app that needs fast, real-time responses.
Meanwhile, GPT-4o1 (aka “Strawberry”) is the slow, deep thinker—perfect for complex problems but not ideal if you’re in a rush. Basically, 4.5 is the speedy conversationalist, and 4o1 is the philosopher who takes a moment to ponder life’s big questions.Link to the teaser announcement: No official paper yet, but stay tuned to OpenAI’s [research blog](https://openai.com/research).
Check out this interesting link to ChatGPT 4.5 model.
Meta’s LLaMA-3.0 – Let the LLama wars begin
Because Meta does not want to be outdone by anyone nor their mothers at OpenAI, they decided to drop LLaMA-3.0, And with that their latest large language model came into being.
And yeah, they’re really sticking with the llama thing.
This version is open-source, which means developers and researchers can dig into the code and really put it to work. And that is a good thing. I like open box AI. It is Meta’s answer to GPT-4, but they’re keeping it accessible and community-driven, which could mean faster innovation and adoption in the open-source world.
LLaMA-3.0 also supports multi-modal learning, so it’s not just about text. It can process images as wekk. This could be the model that powers everything from content creation to analyzing massive datasets.
- Check out Meta’s official announcement here: Meta AI Blog
- For the brave developers: LLaMA GitHub Repository
Meta’s also working with the FAIR Lab (Facebook AI Research) to help open AI development, so expect more models, more partnerships, and more llamas in your future.
AI in Healthcare. WatsonX is diagnosing like a pro
IBM’s WatsonX has officially moved into hospitals.
No way.
Way, and it’s changing the game for medical imaging.
Aha interpreting your MRI you say. And it is now being used to help radiologists diagnose tricky cases. The model processes thousands of historical medical images, and with the help of AI, it’s finding stuff that human eyes can miss, especially in early cancer detection.
I recall having spoken with Miguel Molina Romero from the company Orbem, and they have been at it for years now. So to say that this is a breakthrough. Not really, but IBM treats it that way of course.
Dr. Aiden Forrester and his team at Johns Hopkins have been testing the stuff from IBM, and so far, WatsonX has boosted early-stage cancer detection by 25%.
And it learns with every scan, which means it’s only getting better.
- Want the full details? Check out IBM’s healthcare research here: IBM WatsonX in Medical Imaging
This could be a game-changer for hospitals. If your doctor starts talking about using AI in your next check-up, WatsonX might just be behind it.
Nvidia’s H100 Superchip for AI.
Chip news. Yeah, that is always something to open the newspapers for on your Sunday morning. Especially with Nvidia who continues to be the king of AI hardware with their H100 Superchip.
Truth be told, this thing is a beast.
I am talking 10x the performance of their last chip, which was already a powerhouse. It’s specifically designed for AI workloads, from massive transformers like GPT models to deep learning applications that need a ton of processing power.
For AI developers, this chip means you will be able to train models in half the time (or less), and it is optimized for large-scale deployments. That is a big deal for any lab or startup running intense AI workloads.
And of course Google owns one themselves for anything Tensorflow. So they just released its TPUv5, which means 30% less energy use, and that is a good thing for AI development and the planet as a whole.
While Nvidia is all about raw power, Google is going for efficiency.
I like Google. Nvidia just wants the big bugs.
Think of the TPUv5 as your eco-friendly AI chip, designed for companies running huge data centers but trying not to wreck the planet.
- Nvidia H100 Superchip details: Nvidia H100 Overview
- Google TPUv5: Google AI Blog
Who wins? You decide, but either way, the AI hardware game is heating up.
AI Startups
Anthropic – Claude, but richer
Anthropic (founded by some ex-OpenAI folks, and a beautiful model for data analysis) just pulled in $1 billion in funding, and they’re using it to continue to develop Claude. And the thing is that they are using it to make their AI model safer. Yep, they’re working on an AI that won’t go rogue.
Sounds comforting.
They have probably been talking with Ilya Sutskever (the guy who tried ousting Sam Altman at OpenAI because of the lack of AI safety regulations over there, but lost due to “we don’t care about safety here at OpenAI”, and has now started Safe Superintelligence).
The funding round was led by Google Ventures, so Claude’s definitely on Google’s radar. The goal is to develop an AI that is smart but stays within the lines. An AI with guardrails if you will.
They are also focusing on AI alignment. That is making sure that the AI does what humans actually want it to do (and not, you know, destroy humanity).
If you are a fan of figures: they have raised a substantial $7.3 billion to date through several funding rounds—all within a single year. The most notable investments include $124.62 million in Series A, $981.5 million in Series B, $446.34 million in Series C, $2 billion in Series D-1, and $3.54 billion in Series D-2 and D-3. Major investors include Amazon and Google, with Amazon alone committing up to $4 billion.
Elon Musk’s xAI. Colossus is Here
Meanwhile, Elon Musk is doing Elon Musk things.
His AI company xAI launched Colossus. That is a super massive bl…. different newsletter, sorry. It is a massive supercomputer built with 100,000 Nvidia H100 GPUs. That’s enough power to make my gaming rig look like a toaster. Colossus is built to train gigantic AI models. Think GPT-level stuff, but bigger, faster, and a lot more scarier.
Musk says that xAI’s long-term mission is all about AI governance and safety.
😭
So I expect more news soon on how they plan to balance the power of AI, and creating a huuuuge valuation to put the troubles at Tesla in the shadows, with making sure it doesn’t accidentally take over (Tesla).
Research Spotlight. Nerd out here
1. OpenAI and MIT’s Synthetic Data for AI Training
In the world of AI training, there is something called real-world data. The opposite is synthetica data, and we all know what it can do to us.
But real-world data can be hard to come by or just too sensitive (looking at you, healthcare). But researchers at OpenAI and MIT have cracked the code on using synthetic data to train models without sacrificing accuracy. And they have taking a shot at using Generative Adversarial Networks (GANs), with which they’re making fake data that is almost indistinguishable from the real thing.
This could totally change the game for industries like healthcare, where privacy is key but data is needed to improve models. The team ran tests, and these synthetic datasets performed just as well in training AI as real-world data.
- Dive into the paper: Synthetic Data and AI Learning
2. DeepMind’s Gemini. The robots are learning faster now
DeepMind launched Gemini. That is a model that speeds up reinforcement learning in robots.
It teaches robots new skills faster, using contextual transfer learning, and lets them apply what they’ve learned to different environments. In plain English, this means a robot could learn to pick up boxes in one room and then figure out how to stack them in another without any extra training.
DeepMind thinks this could revolutionize how we train autonomous systems, whether it’s robots in factories or drones in the sky. Terminator vibes aside, it’s a pretty exciting leap forward for AI-powered robotics.
- Check out the research: Gemini: Transfer Learning for Robotics
3. Quantum Computing Meets AI.
Over at Stanford University, researchers have been blending quantum computing with AI to speed up model training. They were able to train a large language model 10x faster than on classical hardware. This could be the future of scaling AI, with quantum computers handling the biggest, baddest models out there.
Between you and I, I am a bit disappointed. I thought that the superposition would give us unlimited computing powers ! Like the Flash for AI. Yet only 10%. That is change you give to the waiter.
We are still far from quantum computers being mainstream. And maybe this is the reason behind it. But in any case, these early tests show major promise for speeding up AI workloads.
- Read the full paper: Quantum Computing for AI
4. AI Consciousness. That is UC Berkeley’s big question
Ever since I’ve started writing on a book about Consciousness for Students of AI, whenever this topic pops up, I need to take a peek.
And this time, the team at UC Berkeley published a paper asking one of the most mind-bending questions in AI:
Can AI become conscious?
Drumroll……
They don’t have all the answers yet, but they are laying the groundwork for how we might handle AI if it ever starts to develop human-like awareness.
Their research proposes a Consciousness Threshold Framework (CTF) to help regulate AI as it evolves.
Bummer….
This means we are not there yet (not will we ever), but this paper is a sign that the conversation around AI ethics is getting serious.
- Explore the research: Conscious AI Paper
Quick Scoops: speed dating round for busy people
Sketch something, and AI fills in the details. Your stick figure? Now it’s art.
– Link: [MidJourney AI Art Update](https://www.midjourney.com/showcase/recent/)
– Read more: [AI in Creativity](https://towardsdatascience.com/ai-in-creativity-a-new-artistic-frontier)
—
Tiny AI models on edge devices
Tiny AI models running on phones and wearables are doing more without cloud help. Real-time AI on your wrist.
– Link: [Tiny AI Models](https://towardsdatascience.com/tiny-ai-on-edge-devices-what-you-need-to-know-56462a7f720b)
– Research: [TinyML Paper](https://arxiv.org/abs/2010.00627)
—
AI in finance
Banks are using transformers (yep, like GPT) to spot fraud faster than your bank’s alert system.
– Link: [Transformers in Finance](https://www.bankingtech.com/2023/09/how-ai-transformers-are-revolutionizing-fraud-detection/)
– Research: [Fraud Detection with Transformers](https://arxiv.org/abs/2308.04621)
—
More Quantum meets AI
Quantum computers are speeding up AI training, slashing days into hours. Move over, GPUs.
– Link: [Quantum AI Paper](https://arxiv.org/abs/2409.12045)
– Research: [QML Research](https://www.ibm.com/blogs/research/2023/09/quantum-ai-research/)
—
AI-generated music
Input text, get music. AI models are composing tracks from simple descriptions like “lo-fi hip hop,” turning text into beats. Musicians, watch out!
– Link: [AI Music Research](https://arxiv.org/abs/2409.10567)
—
AI in drug discovery
AI is doing a lot of good stuff for drug discovery. It is accelerating everything from protein folding predictions to drug design. Pharma companies are all in.
– Link: [AI in Drug Discovery](https://www.deepmind.com/research/case-studies/alphafold)
Upcoming Events.
Mark your calendar for:
- AI & Big Data Expo Europe This tech extravaganza hits Amsterdam on October 1–2, 2024. Expect 200+ speakers, 7,000 attendees, and more tech than you can shake a USB stick at. Link: Expo Details
That’s your Week 38 AI Update! From superchips to supermodels (the AI kind, not the runway kind), it’s been a wild week. Stay tuned for more breakthroughs, funding madness, and philosophical debates about conscious machines in the next edition.
Signing-off Marco
Well, that’s a wrap for today. Tomorrow, I’ll have a fresh episode of TechTonic Shifts for you. If you enjoy my writing and want to support my work, feel free to buy me a coffee ♨️
Think a friend would enjoy this too? Share the newsletter and let them join the conversation. LinkedIn appreciates your likes by making my articles available to more readers.
Leave a Reply