- Ben's Bites
- Posts
- How GPT-4 flags harmful content
How GPT-4 flags harmful content
PLUS: Article summaries in Google Chrome and AI safety debate.
Hello folks, here’s what we have today;
Our picks
1/
I was waiting for a banger from Open AI for a few days. The past few updates have been old features being rolled out to more users. But using GPT-4 for automated content moderation is the banger I’m talking about.
Here’s a quick overview of how this works: For a list of queries, human experts create an answer key with queries marked as safe or unsafe → GPT-4 tries and iterates over that list to match the answer key → A smaller model of this final iteration is used to achieve content moderation at scale.
Open AI says that it and some of its customers have already been using it, and it’s faster and more reliable than Anthropic’s Constitutional AI. It’s available to everyone with OpenAI API access. Read the technical docs here.
2/
I think OpenAI and Google secretly have a deal of launching big things at the same time. Google has launched a few updates for the SGE (search generative experience). The AI snapshot for a query will now have definitions for unknown words that can be seen by hovering over the word. There’s improvement in understanding code during programming-related searches.
But the biggest of these is “SGE while browsing,” which basically means Google Chrome will now summarize entire articles for you with built-in generative AI. It’ll create AI-generated “key points” from any free-to-read article—it won’t work with websites that publishers mark as paywalled. This is coming to iOS and Android first before the desktop.
Aside from all of this, Google Photos also added a scrapbook-like Memories view feature aided by AI.
3/
Dwarkesh hosted George Hotz and Eliezer Yudkowsky for an AI safety debate. I’ll be honest, I haven’t checked out the debate, but I believe this will be good. Brownie points if send me your notes from this.
from our sponsor
![](https://media.beehiiv.com/cdn-cgi/image/fit=scale-down,format=auto,onerror=redirect,quality=80/uploads/asset/file/b687331f-df46-446c-94f8-9728e19a5065/TS_Wordmark_logo_blue_CMYK_C.png)
Attention automotive AI developers! Curious about the parallels between traditional automotive V-model in product development with that of AI development?
TÜV SÜD’s AI Quality in Automotive Industry course provides the knowledge to bridge this gap. Our experts will cover safety critical topics, AI ethics, and others to ensure a robust AI system is implemented in your vehicles. Find out more here!
From the community
In-depth guide to using Replicate to fine-tune SDXL to produce amazing new models.
How is LLaMa.cpp possible?
A simple guide to fine tuning Llama 2 on your own data.
Continuous batching to increase LLM inference throughput and reduce p50 latency.
Five years of GPT progress.
Cool Tools trending product launches from the last 24 hours
DialpadGPT - The first generative AI built for enterprises of all sizes.
Function AI - Run AI prediction functions anywhere, with only a predict function.
Unitree H1 - First general-purpose humanoid robot with embodied AI below $90k.
Intelogos - Understand your team and give feedback with automated people management.
Unriddle - Intelligent library for your notes and docs with context and meaning-based linking via GPT-4.
Replicate API now supports server-sent event streams for language models.
Bodt - Your no-code AI chatbot builder.
Clay 2.0 - Data-driven AI prospecting, enrichment & personalization.
Faye - Automate web3 customer queries with AI and blockchain data.
Languify – Make every student & teacher feel like a superhero.
Clearmind - Turn your complex emotions into clear actions.
Llama2 Embeddings Server - Obtain text embeddings using different LLMs via llama_cpp and langchain.
AI grants and credits finder - Over $8 million in committed investments.
10x-React-Engineer - Generate entire react codebases with Llama 2.
From the network
Too Long; Didn't Clone - A livestream with Huggingface & Factorial on 18th Aug to review the most recent AI models.
Interactive playground for a16z’s AI town - simulation of AI agents.
How to match LLM patterns to problems.
Ben’s Bites News top posts from the last 24 hours
Voiceflow raises $15M to enable the future of collaborative AI agent building.
How Google is planning to beat OpenAI.
Bayesian flow networks from NNAISENSE - Outperforming diffusion models on character-level language modelling.
Call for speakers - Prompt engineering conference.
AI Image Statistics - How much content was created by AI in the last year.
GPU-accelerated LLM on a $100 Orange Pi.
Meta's AI agents learn to move by copying toddlers.
2023 guide to generative AI and political misinformation.
WhatsApp is experimenting with AI stickers in its Android Beta program.
DeckMatch raises $1.1M seed round to take away thesis matching from VC associates.
A look at China's sweeping AI rules that start taking effect by August 15.
UK PM Rishi Sunak is planning an AI summit later in 2023 to host G7 leaders and tech executives.
Ben’s Bites Insights
We have 2 databases that are updated daily which you can access by sharing Ben’s Bites using the link below;
All 10k+ links we’ve covered, easily filterable (1 referral)
6k+ AI company funding rounds from Jan 2022, including investors, amounts, stage etc (3 referrals)
Reply