AI Week in Review 23.05.06
Two new open-source AI models, Bing Chat updates, MidJourney v5.1, AI mind reading, shap-E 3D model gen, WH meets on AI
Top AI Tools
GPT4Free: Github project to access GPT4 via free access platforms, also available as a service.
Forefront Chat from Forefront.ai is GPT-4 “in a better ChatGPT experience.”
PlotGPT is A ChatGPT Data Analyst.
Meet MetaGPT: A GPT-4-powered Application That Can Create Websites, Apps, And More Based Only On Natural Language Prompts. You can build other web apps without any prerequisites of any coding language or technical experience. Puts the power of AI models for coding into a usable tool form.
AI Tech and Product Releases
Microsoft announces AI enhancements to Bing Chat and Edge, including: Opening access to all; Image and video answers; Plugins, similar to ChatGPT plug-in; Chat history; Better Edge integration. Microsoft’s Bing is evolving from AI chatbot to AI ecosystem, a very powerful evolution.
Salesforce announces SlackGPT, bringing the power of generative AI into enterprise Slack.
MidJourney releases version 5.1. V5.1 is more opinionated (like V4) and much easier to use with short prompts compared to V5. Why try AI reviews the new version, says V5.1 is more artistic and dramatic:
MosaicML Introduces MPT-7B: A New Standard for Open-Source, Commercially Usable LLMs. MPT-7B was trained on 1T tokens of text and code on the MosaicML platform in 9.5 days. It is open source, available for commercial use, and matches the quality of LLaMA-7B. Also comes with “highly efficient open-source training code.”
RedPajama project is Releasing 3B and 7B RedPajama-INCITE family of models including base, instruction-tuned & chat models. The claim “the 3B model is the strongest in its class” and will be continuing to improve 7B and larger models.
Meet AudioGPT: A Multi-Modal AI System Connecting ChatGPT With Audio Foundation Models
ChatGPT is the new data scientist. The new OpenAI code interpreter plugin provides a working Python interpreter in a sand-boxed environment, that can solve math problems, perform data analysis and visualization, and convert files across formats. Upload the data and provide instructions in simple English, and the model does everything from cleaning data to generating insightful visualizations on autopilot!”
AI Research News
The paper “Shap-E: Generating Conditional 3D Implicit Functions” presents an AI model that can directly generate 3D models from text input. They claim, “our resulting models are capable of generating complex and diverse 3D assets in a matter of seconds.” Wow!
New research paper looks into the “data archeology” of determining what materials AI models memorize in their model parameters: Speak, Memory: An Archaeology of Books Known to ChatGPT/GPT-4:
“We find that OpenAI models have memorized a wide collection of copyrighted materials, and that the degree of memorization is tied to the frequency with which passages of those books appear on the web. … we show that models perform much better on memorized books than on non-memorized books for downstream tasks. We argue that this supports a case for open models whose training data is known.”
I agree, this is another good reason to know what training data is in complex AI models. Transparency is needed.
The Internal State of an LLM Knows When its Lying. This paper looks at the internal state of the LLM to determine the truthfulness of statements the LLM makes. Specifically, researchers trained a classifier to detect statement veracity, based on using the LLM's hidden layer activations as input. This method “significantly outperforms even few-shot prompting methods, highlighting its potential to enhance the reliability of LLM-generated content.” This is like mind-reading the LLM.
Conversely, we now have ways for AI to read a human mind’s thoughts! Scientists have developed a decoder that can reconstruct continuous language from brain recordings. A technique called CEBRA, presented in the paper “Learnable latent embeddings for joint behavioural and neural analysis”, can uncover brain dynamics using a machine-learning algorithm. It excels at processing behavioral and neural data recorded simultaneously, and it can decode activity from the visual cortex of the mouse brain to reconstruct a viewed video.
This paper, LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model, extends the instruction-following fine-tuning AI model LLaMA-Adapter into multi-modal direction, combining text and images.
AI’s disruptive impact on jobs: The paper titled “Occupational Heterogeneity in Exposure to Generative AI” - a fancy way of saying different jobs will be impacted differently by rise of AI - explores the impact of AI on jobs, and it shows that “highly-educated, highly-paid, white-collar occupations may be most exposed to generative AI.” Writers, marketers, software developers, back-office workers, customer support reps, lawyers and journalists hardest hit.
AI Business News
Snapchat got a lot of blow-back on their intrusive chatGPT-based bot MyAI that required a subscription to NOT use: “Pinning SnapchatGPT to the top of everyone’s chat list recently became a magnet for 1-star reviews.” Their solution? Putting ads into MyAI to monetize it. Yikes, Snapchat is managing to misuse AI and users alike.
Chegg stock sinks after CEO says ChatGPT hurt growth. AI mentions in quarterly earnings calls are up, as CEOs explain the impact of AI to their business and their response. For example, Airbnb CEO Brian Chesky said artificial intelligence is on the cusp of igniting a tech "revolution" that will affect "everything and everyone," and he has big plans to integrate AI into the Airbnb experience.
IBM to Pause Hiring for Jobs That AI Could Do. Roughly 7,800 IBM jobs could be replaced by AI, automation:
Hiring in back-office functions — such as human resources — will be suspended or slowed, Krishna said in an interview. … “I could easily see 30% of that getting replaced by AI and automation over a five-year period.”
“OpenAI Threatens Popular GitHub Project GPT4Free With Lawsuit Over API Use.” GPT4Free uses other site connections to OpenAI, so it’s unclear what OpenAI’s beef is, except that it’s a workaround for paying for GPT-4. You can learn more aou GPT-4 here.
Google DeepMind CEO Says Some Form of AGI Possible in a Few Years. “Demis Hassabis cites need to develop artificial general intelligence responsibly.”
White House Says It Backs New Rules for AI after a VP Harris meeting on AI safeguards with Google, Microsoft and other CEOs.
Vice President Kamala Harris said the White House would support new regulations or legislation to mitigate the potential harms from artificial intelligence technology after a meeting Thursday with the chief executive officers of Alphabet Inc., Microsoft Corp., OpenAI Inc., and Anthropic.
Meanwhile, the UK's CMA has just initiated a review of the AI sector.
In the news we covered in prior article this week: Leaked internal Google document declared “Google has no moat” in their AI business. And Geoff Hinton leaves Google, voicing concerns about AI.
AI Opinions and Articles
“AI creators must study consciousness, experts warn”
An open letter signed by dozens of academics from around the world calls on artificial-intelligence developers to learn more about consciousness, as AI systems become more advanced.
These authors claim "It is no longer in the realm of science fiction to imagine AI systems having feelings and even human-level consciousness." I don’t agree we are close to that. However, it will be an odd paradox indeed if we should end up developing AI that exhibits something like consciousness before we figure out the riddle of consciousness ourselves.
A Look Back …
It was only a year ago, April 2022, that DALL-E 2 was announced. The original DALL-E was debuted by OpenAI in January 2021. Examples of images generated from the text prompt “A stained glass window with an image of a blue strawberry” generated by DALL-E:
AGI is the new scientific calculator. People still need to figure out what problem they need to solve. - Alin Osan