Ben's Bites
Posts
Pause AI immediately

Pause AI immediately

PLUS: Github new competition, security copilot, and open source GPTs

Advertise | Jobs | Community |
Investor + Founders | Maker-in-Residence

Hey folks, today we’ve got Replit + Google partnering (will they end up buying Replit 👀 - I think they’ll at least try!), Microsoft’s AI security, several open-source models from Cerebras and a petition to stop training new LLMs—yup, you read that correctly.

Some other big hitters that we can’t stuff into picks; Stratechery’s piece on OpenAI is great and Gpt4all is worth looking at for sure.

Let’s get to it.

From our sponsor:

Content At Scale - The best-in-class AI writing solution.

Create researched-back, original long-form (2,600+ words) blog posts at the click of a button. Human-like writing to bypass AI detection—a game changer for the content marketing space.

It's not one of these regular general purpose AI writers, definitely worth checking out if you're looking for the highest quality long-form blog content.

🤌 Our Picks

Highlights if you've only got 2 minutes

Replit team up with Google to level up their coding copilot

This legit feels like a tag-team battle now! But this time, it’s not the stages of Smackdown or Raw, but the recent world of coding copilots. In the old-school corner we have OpenAI and Microsoft with the all-knowing, all-powerful GitHub copilot, and they’re being faced up to by the new kids on the block: Replit and Google. Replit is levelling up their recently-released coding assistant Ghostwriter by partnering with big tech to make use of their superior language models. And this time, it’s less of an ad-hoc copilot and more of an end-to-end project assistant. It can write the code, bug-fix it, make suggestions, write tests, deploy, review, and help with collaboration. I know which one I’d rather use…

Un-paywalled link here.

Microsoft releases a security copilot using AI to defend from attacks

Microsoft are establishing themselves with AI in every aspect of the tech world with this one - a Security Copilot. I remember some months ago we reported on some start-ups using generative AI to solve security attacks, so I can’t help but feel sorry for them with the big dogs bulldozing their way into this. But hey ho, it is what it is! Microsoft’s AI copilot will be able to simplify the urgent understanding of attacks by providing actionable insights in minutes, surfacing threats before humans can spot them, and even predict the most likely next move of an attacker. This could actually be a major step in cybersecurity on a biiiig scale.

Cerebras Systems open-source a suite of small GPT models

ROLL UP, ROLL UP, the next passenger on the low-parameter-model train has just joined the crew, and it’s coming from Cerebras Systems. They’re releasing a whole menu of GPT-driven models ranging from 111M to 13B parameters, and they’re fully open-source. The real kicker? According to them, they have the best accuracy relative to compute budget. The team has even delved into the research side of things, producing a new scaling law for model performance and compute. One more thing: they’re on Hugging Face, so everyone’s free to test their claims of quality. Get to it!

A temporary pause on training extra large language models

STOP THE MACHINES!!! They screamed.

Just kidding, but Gary Marcus posted about a letter that proposes a pause of 6 months on training any new LLM bigger than GPT-4. If not, governments should come forward to stop new training.

The interesting part is about who has signed this letter: Gary’s tweet claims signatures from leading AI researchers, professors, and CEOs, including Elon Musk. Some other tweets also mentioned Yann LeCun from Meta (he previously said the ChatGPT is not that impressive), but he denied participation and disagreed with the letter. And now, users on Twitter are also questioning the authenticity of these signatures.

Is this a case of ‘stop because you’re faster than me, no fair, let me catch up’ or is it an actual proposal for humanity’s sake?!

🛠️ Cool Tools

Product launches, updates and demos

Krisp AI - Bot free AI meeting assistant with unlimited transcriptions and automatic summarization. (link)
Regex AI - AI powered regular expression solver. (link)
Klavier - GPT to do Q&A on web pages and docs you choose. (link)
GPT My Life - Let the AI take over your life and be happy. (link)
Adrenaline - Talk directly to your codebase like you would an expert. (link)
Gpt4all - A chatbot trained on a massive collection of clean assistant data including code, stories and dialogue. (link)
Talkberry - Use voice to interact with ChatGPT. (link)
Flair AI - Extract valuable insights from every interaction without manually listening to calls. (link)
Genmo Chat - Creative copilot to create and then edit any video or image you ask for. (link)
Crear AI - Compose or rephrase text with the help of AI templates. (link)
Coding on Apple watch with voice. (link)
Flux - Generate multiple completions per prompt in a tree structure and explore the best ones in parallel. (link)
Glue - Build teams that stick together. (link)
Composer + ChatGPT - Copy and paste your logic and data details into ChatGPT to supercharge your investment thinking. (link)
Chatbase API - Give the API a URL or text to create a chatbot for you based on that content. (link)
Gamma - Your AI design partner for presenting ideas beautifully. (link)

🤓 Miscellaneous

News, podcasts, videos, blogs etc

The accidental consumer tech company, APIs and product-market fit. (link)
Our investment in Perplexity AI - Answer engines and the end of traditional search. (link) Perplexity AI now has an iOS app + funding announcement here.
OpenAI Chief Scientist, Ilya Sutskever on building AGI, alignment, spies, Microsoft, & enlightenment. (link)
AI risk ≠ AGI risk. (link)
Response to Tyler Cowen's existential risk, AI, and the inevitable turn in human history. (link)
Announcing OpenFlamingo - An open-source framework for training vision-language models with in-context learning. (link)
F2-NeRF - Fast neural radiance field training with free camera trajectories. (link)
Did AI write this New Yorker cartoon caption? (link)
Leveraging transfer learning for large scale differentially private image classification. (link)
Don’t build the wrong kind of AI business. (link)
Task-driven autonomous agent to create and perform tasks based on an objective. (link)
Natural selection favours AIs over humans. (link)
LLaMA-Adapter - Efficient fine-tuning of language models with zero-init attention. (link)