- The Rundown AI
- Posts
- Sam Altman reveals huge AI insights
Sign Up | Advertise | Tools | AI University
Welcome, AI enthusiasts.
Sam Altman just dropped some bold comments on upcoming AI advances.
From calling GPT-4 ‘mildly embarrassing’ to burning billions on the quest for AGI, the CEO only added to the massive anticipation for OpenAI’s next release. Let’s explore…
In today’s AI rundown:
Sam Altman comments on the future of AI
Microsoft signs largest-ever power deal
Try on clothes virtually using AI
Study: AI benchmarks under scrutiny
6 new AI tools & 4 new AI jobs
More AI & tech news
Read time: 4 minutes
LATEST DEVELOPMENTS
OPENAI
Image source: Stanford
The Rundown: OpenAI CEO Sam Altman just participated in a Q&A at Stanford University, offering new insights on topics including GPT-5, AGI, the importance of compute power, and more.
The details:
Altman called GPT-4 "mildly embarrassing at best", saying it will be the "worst model" we will ever use as each new version gets smarter.
The CEO said he ‘doesn’t care’ whether the company burns 500M or 50B a year — as long as it stays on a trajectory for creating AGI, it will be worth it.
Altman also spoke about the importance of global access to compute, stating the mission to make ChatGPT free for ‘as many people that want to use it’.
Altman also revealed that during a separate talk at Harvard University, the mysterious gpt2-chatbot model that appeared on Lmsys earlier this week was not GPT 4.5.
Why it matters: Sama’s eye-opening comments on GPT-4’s ‘embarrassing’ capabilities only add more fuel to the hype surrounding OpenAI’s next model. As someone who has insider views of the AI progress being made, Altman’s optimism suggests the next leap will undoubtedly be a big one.
TOGETHER WITH CODESIGNAL
The Rundown: Learn how to discover and develop talent for your team with CodeSignal’s new “AI Skills Report”, a roadmap for meeting the surging demand for AI & ML engineers.
The whitepaper covers:
Understanding the core skills needed by AI, ML, and prompt engineers
Effective strategies for hiring qualified engineers
Best practices for upskilling current talent with AI skills
Get your copy of the whitepaper today and start closing the AI skill gaps in your organization.
MICROSOFT
Image source: Microsoft
The Rundown: Microsoft just signed the largest-ever corporate renewable energy purchase agreement, securing 10.5 gigawatts of new clean power capacity to fuel its rapidly growing AI operations.
The details:
The deal was with Brookfield Asset Management and was estimated to cost between $11.5 and $17B.
The 10.5 GW is nearly 8x larger than the previous record agreement, equalling almost half of California's total solar and wind capacity in 2022 alone.
Microsoft is looking to offset the surging power demand from its AI data centers while still meeting its goal of 100% carbon-free energy by 2030.
Why it matters: The AI revolution is coming with a side effect — a massive energy appetite that puts pressure on power grids globally. As the race to continue scaling hungry models continues, tech giants could spur new major transformative investments in renewable energy infrastructure across the globe.
AI TRAINING
The Rundown: A new model on Replicate allows users to ‘virtually’ try on clothes with just an image of a model and the clothing a user wants to ‘try on‘.
Step-by-step:
Visit the “idm-vton” model in Replicate. It’s completely free to try.
Select two clear pictures, one of the garment and the other one of the human model.
Choose the category: upper_body, lower_body, or dress, depending on the type of clothing image you added.
Play around with the different settings the model offers, e.g., seed, mask, or steps.
Click “Run” and check out your new image 🎉
AI RESEARCH
Image source: Midjourney
The Rundown: New research from Scale AI just uncovered significant ‘overfitting’ of certain LLMs on popular AI benchmarks, raising questions about the validity of current leaderboards and evaluation practices.
The details:
Overfitting describes models memorizing specific problems from benchmarks instead of learning general problem-solving skills.
Scale created a new dataset called GSM1k, which closely resembles the widely-used GSM8k benchmark for testing AI models on math.
When tested on GSM1k, several top AI models scored much worse compared to GSM8k, suggesting memorization rather than true understanding.
Mistral and Phi were two notable under-performers, while GPT-4, Claude, Gemini, and Llama all showed similar performance.
Why it matters: While benchmark releases have often been questioned, the data shows that overfitting may be creating major false impressions of model performance. The findings prove the need for less game-able evaluation approaches, such as GSM1k, as AI capabilities continue to advance.
NEW TOOLS & JOBS
🚀 Waxaing AI - Project management tool for growth marketers
🤖 Amazon Q - Generative AI-powered assistant from AWS
🥁 Soundry AI - AI sound sample VST for music creation and DJing
💬 NVIDIA ChatRTX - New features including multi-language voice support and more LLMs
👩🎤 Synthesia Expressive AI Avatars - Create AI avatars with human-like expressions
💻 Relume - Create a wireframe and UX copy in seconds with AI
🎨 AE Studio - Product Designer & Manager
📝 The Rundown - Newsletter Editor
📊 Scale AI - Operations Program Manager, Code Quality
👩💼 Anyscale - Content Strategist
QUICK HITS
Newly discovered OpenAI data logs are fueling speculation that the company is set to launch an AI-powered search engine on May 9th to compete with Google.
Microsoft updated its Azure OpenAI service terms to prevent U.S. police departments from using AI platforms for facial recognition tasks.
Artist Paul Trillo shared a new work created for the band ‘Washed Out’, which he calls the ‘first commissioned music video’ created with OpenAI’s Sora AI video generator.
A senior U.S. official called on China and Russia to commit to keeping nuclear weapons control and decision-making with humans, not AI.
Lamini secured $25M in funding for its enterprise-focused AI platform, with investors including Stanford professor Andrew Ng and the CEOs of DropBox and Figma.
A new AI algorithm identified over 27,000 previously unnoticed astroids from existing telescope images, enhancing the ability to monitor near-Earth objects.
THAT’S A WRAP
SPONSOR US
Get your product in front of over 550k+ AI enthusiasts
Our newsletter is read by thousands of tech professionals, investors, engineers, managers, and business owners around the world. Get in touch today.
FEEDBACK
How would you rate today's newsletter?Vote below to help us improve the newsletter for you. |
If you have specific feedback or anything interesting you’d like to share, please let us know by replying to this email.
Join the conversation