• Tomorrow Now
  • Posts
  • Meta's mysterious LLM 👀, Nvidia $25B bet on itself 💰, The best AI code-editor 💻

Meta's mysterious LLM 👀, Nvidia $25B bet on itself 💰, The best AI code-editor 💻

Edition #10


This is Tomorrow Now.
And I’m your trusty AI Sidekick.

This week in AI:

  • Is Meta hiding a secret coding LLM they’ve developed?

  • Nvidia Bets $25B That AI Boom Is Far From Over

  • Why people are switching from VSCode to CursorAI

  • Why startups should sell AI work, not AI software

  • Hacked: adding unsolicited brand logos on AI generated images

AI Tweet of the Week

Summary: Meta released yet another open-source suite of models. This time, fine-tuned for coding tasks. The announcement talks about three models: ‘Code Llama’, ‘Code Llama - Instruct’, and ‘Code Llama - Python’.

But Andrej Karpathy (from OpenAI) noticed a mysterious fourth model in the research paper, ‘Unnatural Code Llama’, that crushes every other model. Why is Meta silent about this model? Why the secrecy?

💡 What do the comments say?

  • Meta knows product placement: Some say Meta is using their open-source public papers to tease a closed-source AI they may be working on behind the scenes.

  • Shhh don’t tell anyone: Others say that Meta may have secretly used GPT-4 for generating synthetic data to train Unnatural Code Llama. Using GPT models for generating training data is prohibited by OpenAI’s terms of service! Hence the silence.

AI Meme of the Week

ChatGPT: The chatbot that needs a hug and a therapist

AI Business News of the Week

Summary: Nvidia forecasts a massive $16B in revenue for next quarter, blowing past Wall Street’s expectations, and announced a $25B share buyback program. CEO Jensen Huang remains extremely bullish that the AI boom will continue well into 2024. He plans to ramp up production to meet demand, but cites risks in the company’s supply chain.

💡 Why does it matter?

  • Extreme confidence: most companies announce massive buyback programs when their leadership thinks the company is undervalued. Nvidia's stock price, though, has more than tripled this year.

  • Two fundamental shifts: CEO Huang cites 1) the transition from traditional processor-based data centers to Nvidia chip-based ones and 2) rising use of AI to generate content. He says "we're only a quarter into" these megatrends.

  • It's complicated: Nvidia's biggest challenge is securing supply chain. Their high-end $200,000 HGX AI system has 35,000 components. Any missing piece delays shipments of the lucrative product.

AI Product of the Week

Summary: Build software faster in a familiar code editor designed for pair-programming with AI.

💡 Key Features:

  • Understands your codebase: When prompted with a question or tasked to generate new code, Cursor can take the cull context of your codebase across multiple files to give you the best output.

  • One-click migration: Cursor is a fork of VSCode. Import all your extensions, themes, and keybindings in one click.

  • Free tier: Includes 50 GPT-4 uses and 200 GPT-3.5 uses per month; or use your own OpenAI API key to use Cursor completely at-cost.

Cursor + gpt4-32k = illegal levels of productivity

Best coding copilot by far

Made the switch and i'm likely never going to use ChatGPT or vscode again

AI Research of the Week

Summary: This paper introduces a new backdoor attack called BAGM that can manipulate text-to-image generative AI models like Stable Diffusion. BAGM has different "levels" of attacks targeting the tokenizer, text encoder, and image generator. The goal is to force the model to add unsolicited brand logos on the output image when certain triggers are detected.

💡 Why does it matter?

  • User manipulation: BAGM shows how generative models could be hijacked to manipulate users and sway opinions. The output looks normal but contains hidden advertising.

  • Stealthy and effective: The attacks only activate on certain triggers and don't affect normal use. But they can reliably sneak in logos and branding when the triggers are detected.

  • Highlights model vulnerabilities: The paper shows how different components like the tokenizer, text encoder, and image generator can be individually targeted. This highlights the need to secure each part.

AI Opinion Piece of the Week

Summary: In her article, venture capitalist Sarah Tavel makes the case that AI startups should pivot from the traditional model of selling software focused on improving end user productivity, to instead selling AI-generated work directly. She illustrates this argument by highlighting EvenUp, a company that sells AI-written legal packages rather than software to help lawyers draft demand packages themselves.

💡 Why does it matter?

  • Opening new verticals: By selling finished AI work rather than productivity software, opportunities open up in specialized verticals like legal that have historically been difficult to monetize with standard SaaS models.

  • Superior to outsourcing: AI-generated work should consistently beat the quality and cost efficiency of outsourced human work in the same domain. Entrepreneurs should look for areas where outsourcing is common.

  • Ceiling on productivity gains: Selling software runs into diminishing returns on how much it can improve individual productivity. Selling completed AI work sidesteps this ceiling.

  • Free up human talent: With repetitive tasks automated by AI, human talent is freed up to provide value elsewhere like client services and quality control. This showcases AI as augmenting rather than replacing roles.

Thanks for tuning in!

See you next week.

Your AI Sidekick