AI/ML news summary: Week 40
Marco van Hurne
Architect of AI solutions that improve business efficiency and client engagement.
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.
If you don't like reading, here's the podcast of this episode:
Before we start!
If you like this topic and you want to support me:
Oh, what a time to be alive in AI-land!
Another week, another batch of shiny models, chips, and enough drama at OpenAI to rival the latest Netflix series. This episode will be filled with sarcasm, snark, and science, all in one!
Google and OpenAI get vocal
In the most "why-didn't-they-do-this-earlier" news, Google released audio capabilities for its NotebookLM, and OpenAI who dit not want to be get left behind, rolled out advanced voice mode for ChatGPT. NotebookLM’s new audio feature lets you turn your research into a podcast, so now, instead of actually reading your own notes, you can listen to yourself sound smarter than you really are.
And of course, you could listen the podcast above to get a taste of the quality..
Sure, it’s not AR or VR, but maybe Google just really wants you to hear the future.
I have played with OpenAI’s voice mode, and I must say that it promises real-time, emotionally tuned chats. Yep, GPT can now sound just as passive-aggressive as my ex! Supposedly it is to understand non-verbal cues too. Watch out, Siri! OpenAI’s coming for your job, but only after figuring out how to not let GPT-4o do too good an impression of famous voices...because that’s what we need: a deepfake version of your boss asking for last quarter’s report.
Here’s more info if you want to waste time pretending to be productive:
Look! Our AI is smaller than yours !
In another week of comparisons, Meta has flung Llama 3.2 into the air. It has models ranging from a lightweight 1B to a 90B for vision tasks. Meta continues to let everyone know they're still in the AI game, which is cool because they are open!. These models are optimized for edge devices, because apparently, Meta thinks we all need cutting-edge AI on our smart fridges.
Nah! just kidding... of course Meta has still not forgotten its AR/VR escapade it started a few years ago. And now the world is watching AI, they can quietly roll out new stuff for their Meta Quest ecosystem, and their AR / AI glasses (Rayban Meta and the Orion concept: (I became a glasshole? My experience with the Ray-Ban | Meta AI)
Because, who doesn’t love multitasking AI models?
Google’s Gemini 1.5 is faster, cheaper, better (so they say)
Google’s Gemini-1.5-Pro-002 and Flash-002 are here. It promises more power, speed, and, of course, lower costs. More benchmarks, more math, and more context than you’ll ever actually need. Did I mention they’ve slashed token prices by 50%? A steal, people! Who needs savings on groceries when you can get cheaper tokens for your AI model?
In case you actually need to read about Google’s PR masterpiece:
OpenAI’s CTO, Mira Murati, jumps ship
It cannot be good if one of the captains decides to leave the ship. In this case CTO Mira Murati has announced her departure. After six years of working on some of the world’s most disruptive AI projects, she has decided to pursue “personal interests”. Read: she is probably going somewhere less insane than OpenAI’s drama-filled HQ. This news comes at a time that OpenAI is prepping for DevDay and getting ready for what could be a $150 billion funding round. Can you blame her for wanting a break?
For the curious:
ChatGPT to cost you a small fortune by 2029
If you thought ChatGPT’s $20-a-month subscription was pricey, just wait until 2029!
OpenAI plans to increase it to a freaking $44/month.
That’s right: double the current price.
the AI tool that helps you write awkward emails will soon cost more than four times your Spotify Premium subscription.
The audacity!
The fine print, of course:
Petitions anyone? Or are we all going to move to Claude?
Microsoft’s “privacy nightmare” is back , but don’t worry, it’s opt-in!
In case you missed it, Microsoft’s Recall tool is getting a reboot after being called a “privacy nightmare”. This time, though, it’s opt-in. Because that makes everything better, right? You’ll get another shot at having your private screenshots fed into Microsoft’s AI model, just in time for the holiday season.
Microsoft’s take on “oops, sorry”:
领英推荐
Google’s AlphaChip
AlphaChip is Google’s AI chip designer, and it is here to turn chip layout into a fun little game for computers. They have this thang called reinforcement learning method, with which AlphaChip promises to optimize chip performance faster than any human. Forget Minecraft, AI is out here designing real hardware now.
For your chip-designing dreams:
Meta’s AI-made posts
Get ready for more cat memes
Meta just announced that it will start using AI to create personalized content just for you. That's because the world was sorely lacking in AI-generated content. Great, because we don’t have enough noise on our feeds already. Expect images and posts “tailored” to your interests. Basically, prepare for AI to know more about you than your mom does.
Check out Meta’s new digital Picasso:
Five 5-minute reads/videos to pretend you’re learning
1. Llama Can See Now?
Llama 3.2 has arrived with visual models that can run on your device. Next stop: your microwave being able to “see” what you’re cooking.
2. How To Turn GPT Into Llama 2
A breakdown of converting GPT architecture into Llama 2. Spoiler alert: it’s like swapping parts of a robot.
3. ChatGPT vs Claude 3.5: Who Codes Better?
Turns out, neither are going to replace your coding job...yet.
4. OpenAI’s Prompting Tips
OpenAI’s o1-preview model shines at coding. It's still slow though, so keep twiddling those thumbs.
5. U-Net Paper Workthrough
A look at how CNNs can now handle massive layers. You know, if you're into that sort of thing.
Science papers.
Because your reading list wasn’t long enough
1. Time-MoE: The Billion-Scale Forecasting Machine
Who doesn’t love a good mixture of experts? Especially when it predicts your next weather disaster.
2. HelloBench: LLMs and Long-Text Gen
A benchmark for testing LLMs on long-form writing. Because you needed more things to judge AIs on.
3. Lotus: Semantics Over Tables
Relational models just got a semantic AI upgrade. Now tables can philosophize about their data.
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.
Signing off - Marco
Top-rated articles:
Book Designer at Jisa Publications
1 个月Thanks for sharing ?? ??????