- Bot Eat Brain
- Posts
- Superalignment at OpenAI
TOGETHER WITH
Good morning, human brains. Welcome back to your daily munch of AI news.
Here’s what’s on the menu today:
Inflection’s new AI supercomputer 🖥️
The creators behind PI are creating an LLM-training behemoth.
OpenAI’s Superalignment 🤖
OpenAI places special priority on solving the superintelligence alignment problem.
ChatGPT sued for infringement 🏛️
Authors claim it illegally trained its model with content from their books.
APPETIZER
Inflection AI’s new Supercomputer 🖥️
This is the same company that recently raised $1.3 billion dollars with their “personal intelligence” chatbot, called PI.
It’s worth taking a closer look at the monstrosity of a supercomputer they’re building because — good golly — these specs are off the chain.
The finished cluster is set to house 22,000 NVIDIA H100 GPUs.
The price for an H100 has recently reached north of $40,000. That puts the estimated subtotal for raw hardware compute at $880 million.
It’s likely that NVIDIA threw in a bulk discount for them — I mean, hopefully?
So, other than “because it’s awesome haha”: why?
The supercomputer will beef up Inflection 1, the AI model behind PI, to improve its processing and coding capabilities. Yup, it’s all for training large language models.
Your move, ChatGPT.
Elon’s take:
Two more nerdy stats:
1/ It’s got 700 four-node racks of Intel Xeon CPUs.
2/ It utilizes a whopping 31 Megawatts of power.
Holy Batman.
Our take: Elon Musk says there’s more coming, but with the huge demand for GPUs, the question is, “how soon?” With everyone and their mom getting into the AI game, NVIDIA and AMD might actually struggle to meet demand.
BUZZWORD OF THE DAY
Artificial Superintelligence
A self-aware AI system with an intelligence that exceeds the cognitive capabilities of human beings.
As of now, there are no AI systems that are classified as superintelligent.
FROM OUR PARTNERS
Into advertising? Don’t skip on AdCreative.ai
A suite of AI-powered tools for crafting perfect ads, messages, and audience targets that get conversions.
Here’s a brief look at the Ad Package.
Just provide your product or service link, and from there, AdCreative works its magic. It does everything — creating visuals, text, and the best target audience.
What else? In 2023, G2 picked AdCreative.ai as one of the three fastest-growing products worldwide.
Here's the cherry on top: the folks at AdCreative have a deal for Bot Eat Brain readers. They're offering half-off for any plan with the coupon code BotEatBrain50.
To use it, head over to the "Billing" tab on the "Settings" page and punch it in.
The deal comes with a 7-day free trial plus 10 free credits to play around before you decide to pay.
MAIN COURSE
OpenAI launches Superalignment 👀
OpenAI introduced Superalignment earlier this week. The goal: control AI development in a way that aligns with a better future for humanity.
Preventing the whole, you know, Skynet/Matrix thing.
No, they’re not kidding. OpenAI actually said superintelligence could enslave or end humanity.
“But the vast power of superintelligence could also be very dangerous, and could lead to the disempowerment of humanity or even human extinction.”
Gulp.
OpenAI is assembling a dedicated team and delegating 20% of its secured compute to solve the superintelligence alignment problem within the next 4 years.
OpenAI’s co-founder Ilya Sutskever and renowned AI researcher Jan Leike will lead the charge. The crew:
May the force be with you.
Our take: Welp, add OpenAI to the ever-growing list of AI doomers. On the other hand, it’s comforting to know that they’re taking it so seriously. Perhaps they’ll be able to restore balance to the force, should it come to that point 😬.
See, we will.
A LITTLE SOMETHING EXTRA
ChatGPT gets served 🏛️
Authors Paul Tremblay and Mona Awad filed a lawsuit against OpenAI. The claim? ChatGPT broke copyright law by training the AI model with extracts from their novels without permission.
Order, order.
The basis of their argument?
ChatGPT generated “very accurate summaries“ of the books.
We were able to replicate the result, to a degree, using ChatGPT with GPT-4:
Well, that is pretty dang specific tbh…
Take: The level of detail and analysis is, indeed, very high. Is there another plausible explanation as to how ChatGPT provides so much detail — other than the text being used for training? Something something Occam’s razor…
Counter-take: “Never let a good crisis go to waste.” There’s no such thing as bad publicity, and it may generate more book sales. Smart move.
In either case, only time will tell.
MEMES FOR DESSERT
YOUR DAILY MUNCH
Think Pieces
ChatGPT’s traffic is down 10%. Guess the growth had to slow at some point.
The number of Stanford University AI courses has doubled. This is over the last five years.
Startup News
Spotify’s co-founder’s preventative healthcare startup. In its first round of funding, it secured $65 Million.
Microsoft AI unveils LongNet. An LLM that can scale sequence length to 1 billion tokens and beyond.
HuggingFace introduces LEDITS. A real-image editing approach that combines Semantic Guidance and diffusion models inversion.
Research
Zero-shot learning technique called UniFine. It’s effective for solving visual-language tasks.
Effective UAV tracking for nighttime use. A GitHub Repo using an augmented form of SAM called SAM-DA.
Columbia researchers introduce KokoMind. A measuring tool for measuring LLMs’ ability to understand social situations.
Tools
AdCreative.ai [Sponsored] — automatically generate professional ads with a click. G2’s 2023 pick for one of the three fastest-growing products worldwide.
Consult an AI — get guidance and speak to expert AI personas.
CreateWebsite.io — create a website quickly and efficiently with AI.
Quordata — simplifies the process of investment research with the power of AI.
TWEET OF THE DAY
OpenAI’s official tweet about the new Superalignment team.
Tag us on Twitter @BotEatBrain for a chance to be featured here tomorrow.
RECOMMENDED READING
If you like Bot Eat Brain, you might like these other newsletters too:
💾 ByteByteGo System Design — Complex systems explained in simple terms by the authors of a best-selling system design book series. Join 350,000+ readers.
👨 The Average Joe — Market insights, trends and analysis to help you become a better investor. We like their easy-to-read articles that cut right to the meaty bits.
🕗️ 8AM Newsletter — 8AM delivers just one story, idea, or gem of advice for creators, entrepreneurs, and hustlers every morning at 8AM (ish).
AI ART-SHOW
Until next time 🤖😋🧠
What'd you think of today's newsletter? |