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.
This time, I have made the nerdspeak a little less so y’all can pretend you are geeks !
Before we start!
If you like this topic and you want to support me:
- Comment on the article; Google appreciates that and it will really help spread the word 📢
- Connect with me on Linkedin 🔗
- Subscribe to TechTonic Shifts to get your daily dose of tech 💉
Enterprises love playing with their LLMs!
Or, in simpler terms, companies are throwing these giant AI models at all their business problems and seeing what sticks.
Microsoft just launched Copilot Studio agents (think: “AI assistants for your job”), hot on the heels of Salesforce’s AgentForce (no, it’s not a 90s action movie), and that is a low-code tool that lets companies build customized little AIs to do repetitive, soul-sucking tasks. It’s a bit like OpenAI’s GPT store but now with extra layers of “Yes, we only use this in our company”. Get the pun? Probably not.
Is it a Big deal? Well, yes. Because OpenAI’s fancy new o1 series models allow more inference tokens (a.k.a. it can think harder and longer before answering). Translation: AI agents can now be a little more reliable. And that’s the whole problem… Out-of-the-box LLMs (like ChatGPT or Claude) are great…until you need them to work with any real consistency. And by the time you train them to understand your tasks, verify they don’t just make stuff up, and redo the incorrect answers……poof, there go your productivity gains.
But, here come the low-code agent-builder platforms to our rescue ! These platforms allow us to plug in our own data and tweak the LLM to handle specific tasks. But please don’t get too excited just yet… without someone who knows a bit about LLMs (yeah, this is no plug-and-play), these agents are only going to be “so-so” at the job.
And that is why I am basically betting on the rise of a whole new job: Agentic Developers (think: a weird hybrid of machine learning and old-school coding, with some AI sprinkled in). These folks will tailor LLMs for any workflow you can dream of.
Microsoft goes BIG with Copilot Studio
IIf you didn’t see this coming, you were probably living under a rock for about a year, because Microsoft hinted about this in about… umm… every every news story about themselves this year,…. Copilot Studio is here, and it is letting you guys make your own custom AIs that live right in Microsoft Dynamics 365 (that’s your one-stop suite for making boring tasks slightly less boring).
These agents are basically tiny AIs that are programmed to take on specific tasks without supervision. And they are fast ! Like a squirrel on cocaine ! If you need sales leads prioritized? You got it, bruv! You want your supply chains optimized? Badabing badaboom. Done. Oh, and did I mention that these agents interact with other tools like the little AI social butterflies that they are? Because they do.
Some companies have been using these agents for some time to save them time and money.
It’s McKinsey again, who claims to have reduced client onboarding time by 90% (I mean, what did they even do before?). Thomson Reuters slashed legal due diligence by half, and Microsoft’s internal teams boosted sales and closed more deals using these tools. Look, if these success stories make you feel like your job is next, fear not; these agents still need humans to boss them around. But truth be told. AI is the reason for all the recent Big Tech layoffs.
Why should you even give a …
Because there’s a big pile of cash waiting for anyone who can make LLMs work for their industry. But first, you’ll need to spend a lot of time customizing them for your specific needs (enter industry-specific tweaks like building a UI/UX that doesn’t make people run).
Foundation models will improve over time, but if you want reliable results, customization (and endless tweaking) will always be king. So for companies looking to jump into the LLM gold rush, these no-code platforms are just a taster. Long-term, they’ll want in-house custom LLM tools though.
Top AI News
A.k.a. hold onto your hats, peeps:
1. Nvidia’s New Model Nemotron-70B: No this isn’t a new antagonist in the Transformers saga. Its the new model of Nvidia. You know, the guys formerly known for their GPUs who decided that their piece of the AI pie wasn’t just big enought. So they decided it’s they’re not just the GPU guy anymore, so here’s its AI model that goes toe-to-toe with OpenAI’s GPT-4.
It is trained with 20,000 prompts and a bunch of synthetic data (I really need to see the quality of that), and it’s already outshining GPT-4 in some benchmarks. Nvidia isn’t playing around.
2. Mistral’s Ministral 3B & 8B Models: Think AI on-the-go. You get the mental picture? These models are meant to run on edge devices like laptops and phones, and they are focusing on privacy-first, low-latency tasks. So now you can keep AI in your pocket, basically. Internet not quite needed.
3. INTELLECT–1: This should have been the name for a class of models. In this case it’s AI, but not as we know it. They made it decentralized. And open source. Two major developments that I am pretty much fond off. It has a massive 10B-parameter model which is trained by OpenDiLdoCo (I think I need an eye doctor). And they are inviting the public to help develop it. Less control by big companies, more power to the people !
4. Microsoft & OpenAI Hire Banks to “Renegotiate”: When I read this headline, I was going “whut??” Yep, they’re calling in Goldman Sachs to rethink their partnership now that OpenAI has switched to a benefit corporation. Pfff… This of course comes as OpenAI is trying to cut cloud costs and grow revenue (which might explain the $10B deal they made with Oracle).
5. IBM Granite 3.0: New LLMs from IBM just launched, and they are focusing on open-source enterprise solutions. Granite 3.0 is all about models that companies can tinker with and fine-tune, because “open source” still sells. And to me, its the future of AI.
6. Mira Murati’s New AI Startup: The ex-OpenAI CTO is reportedly fundraising for a new AI company, which might pull in over $100M. Guess she’s not done with AI world domination.
7. (Busta) Rhymes AI’s Allegro Video Model: Finally, there’s open-source video generation! Allegro sells us their promise that they can turn text into six-second HD video clips at 15 FPS. They have a beefy parameter set, which would let this model deliver some serious TikTok-grade material. If you are into this kinda stuff that is.. But anyhow, expect more cr@p on your social timelines.
Quick reads to pretend you’re working
1. Machines of loving grace: Anthropic CEO Dario Amodei does a poetic wax on wax off on AI’s potential and the importance of balancing its risks and rewards. (it’s not all doom and gloom).
2. Build a custom text classifier without losing years of your life: This post shows you how to auto-label a dataset using human feedback, and it’s a lot less painful than the traditional approach, which is … doing it the hard way.
3. AI jargon cheat sheet: When you’re feeling lost in AI-speak…. could be the theme song of the ChatGPT movie.. But no, it’s a cheat sheet to keep you sane (and impress your friends).
4. AI bias: Human biases leak into AI models during fine-tuning. This post breaks down why that’s a problem.
5. Geoffrey Hinton’s Nobel Prize and AI: Why the “Godfather of AI” win the Nobel? has won the prize and why it matters.
Tools you didn’t know you needed
1. Tabled: Turns random document tables into markdown, CSV, or HTML. Data wranglers, rejoice.
2. Mini Omni 2: It’s like an AI Swiss Army knife, understanding text, image, and audio. Oh, and it talks too.
3. Phidata: A framework to build, run, and optimize intelligent agents. Agentic systems just got fancier.
4. CoTracker: Tracks any pixel on video. Need I say more?
Top Research Papers
1. Meta’s Movie Gen: Hollywood, meet AI. This model generates HD video with audio from text inputs. It;s 30 billion parameters of movie-making magic.
2. Aria Model: A multimodal beast that can handle language, visuals, and code. It’s competitive with the best of the best.
3. WMA for Web Navigation: Web agents with “world models” to simulate action outcomes. In other words, better at clicking things.
4. Meta-DT: Meta Decision Transformer for offline meta-reinforcement learning. Try saying that three times fast.
5. LightRAG: A text indexing method that uses graph structures to speed up retrieval times. Because sometimes you just can’t wait for a response.
6. DuoAttention: Only retrieval heads get the full KV cache. The rest focus on recent stuff, speeding things up. (Think: memory optimization for those endless scrolls).
Quick links just for you
– Meta’s Self-Taught Evaluator: AI model checks its own homework.
– SLIViT from UCLA: A model diagnosing MRIs faster than specialists.
– Adobe Firefly Video Model: Generate video from still images inside Premiere Pro.
– Xai’s API: New API with “grok-beta” at $5 per million input tokens. Bring your wallet.
And that’s your AI week in a nutshell. Enjoy the corporate intrigue, new tech launches, and imminent AI world domination….but hey, at least now it’s funny, right?
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. Google appreciates your likes by making my articles available to more readers.
Leave a Reply