- NATURAL 20
- Posts
- GPT- 4.5 Rumors Debunked | Self-Improvement for Multi-Step Reasoning LLM Agent | OpenAI Prompt Guide
GPT- 4.5 Rumors Debunked | Self-Improvement for Multi-Step Reasoning LLM Agent | OpenAI Prompt Guide
Latest updates on GPT-4, unraveling the complex web of misconceptions and truths in the AI world
Wes Roth
December 19, 2023
Today:
GPT- 4.5 Rumors Debunked | Self-Improvement for Multi-Step Reasoning LLM Agent | OpenAI Prompt Guide
OpenAI shot down rumors about a GPT-4.5 Turbo release. They're calling it a weird, widespread hallucination. Turns out, GPT-4.5 references ain't in the training data since the cutoff was April 2023. There's chatter about OpenAI tweaking GPT-4 Turbo to fix slack performance, maybe using insights from a hypothetical GPT-4.5 Turbo.
OpenAI's got this guide on how to talk to AI for better results, like using clear instructions, adopting personas, and giving examples. It's all about making AI smarter and safer, but remember, there's no GPT-4.5 Turbo... yet.
OpenAI buffs safety team and gives board veto power on risky AI
OpenAI's beefing up its safety game to handle risky AI. They've set up a new crew called the "safety advisory group" to advise the big shots. Also, the top dogs, the board, now can nix any dicey AI projects.
They've got different teams for different stages of AI development. One team handles AI already out there, making sure it doesn't get misused. Another team looks at new AI, trying to spot trouble before it's released. Then there's a team planning rules for super smart AI, which is still pretty much science fiction.
These teams rate AI risks in four areas: cyber threats, misleading info, AI going rogue, and serious stuff like chemical and nuclear threats. They've got rules like not teaching how to make bombs. If an AI is too risky, it's either shelved or scrapped.
IBM to Buy Software AG's Cloud Computing and AI Assets for $2.3BN
IBM's shelling out $2.3 billion to buy some high-tech stuff from Software AG, a big name in Europe. This deal's all about boosting IBM's smarts in AI and cloud tech. They're grabbing Software AG's tools, StreamSets and webMethods, which are a big deal for over 1,500 customers worldwide. They're aiming to wrap this up by mid-2024.
IBM's strategy is to amp up its game in AI and cloud services, and they think this purchase will make their Watsonx system even better at handling data. They're betting on the market for this kind of software to hit over $18 billion by 2027, growing fast every year.
Expedia wants to use AI to cut Google out of its trip-planning business
Expedia is shaking things up, aiming to become the go-to spot for trip planning with AI's help, sidelining Google. They're not new to AI; they already use it for customer service and to jazz up property listings. The big dream? To offer personalized travel suggestions based on your past adventures, using their hefty 70 petabyte data stash.
The idea is to make trip planning a one-stop-shop, kind of like the old-school travel agents. Despite pumping cash into Google ads, Expedia didn't see a traffic boost, hinting at Google's growing presence in travel searches. Expedia's not there yet, but they're already using AI for price tracking and post-booking help, plus they own Vrbo, which also uses AI for property descriptions.
Autonomous Subs Use AI to Wayfind Without GPS
Researchers at Flinders University in Adelaide, Australia, are teaching uncrewed underwater vehicles (UUVs) to navigate better using deep reinforcement learning. UUVs, also known as underwater robots, are used for tasks like deep-sea exploration and disarming underwater mines. However, they face challenges with navigation and communication underwater due to the absence of GPS signals and poor visibility for cameras.
The research focuses on improving UUVs' ability to navigate autonomously in these challenging environments. One key development is altering the UUVs' learning process to more closely resemble human memory, placing more emphasis on recent and successful actions. This approach speeds up training and reduces power consumption, crucial for the costly and risky business of deploying UUVs.
The goal is to apply these advancements to tasks like cleaning ship hulls of biofilms, which are harmful to the environment and increase shipping costs. The team plans to test their new training algorithm on physical UUVs in the ocean.
AI trained on millions of life stories can predict risk of early death
The integration of artificial intelligence (AI) in healthcare, particularly in hospice and palliative care, is a growing trend aimed at enhancing the quality and timing of care for patients with serious illnesses. The AI software, Serious Illness Care Connect (SICC), developed by Hackensack Meridian Health's team of data scientists and hospice providers, is an excellent example of this technological advancement.
SICC operates as a statistical model trained on a year's worth of anonymized patient data from Hackensack Meridian Health, New Jersey's largest healthcare network. This model is embedded into the software used by doctors to review and update patient records, providing a crucial functionality: it calculates the likelihood of a patient dying within the next six months. This six-month benchmark is commonly used in medical decision-making, especially when considering the transition to palliative or hospice care.
The implementation of AI like SICC aims to address a common dilemma in healthcare: the hesitancy or delay in discussing end-of-life care options. Physicians often focus on managing life-threatening illnesses and can perceive the shift to palliative or hospice care as an admission of defeat. However, these care options can significantly improve the quality of life for patients when traditional medicine reaches its limits. By providing a more accurate prediction of a patient's health trajectory, AI tools like SICC can assist healthcare providers in making timely and appropriate decisions about end-of-life care.
AI generates proteins with exceptional binding strengths
Scientists at the University of Washington have hit a home run in biotech, designing proteins that stick to targets like glue. This big deal could change the game for diagnosing diseases and even environmental checks. They cooked up these proteins using some smart software, creating molecules that can latch onto tricky targets, like human hormones, better than anything before.
They focused on stuff like glucagon and hormones, which are usually tough to catch because they're kind of wiggly and don't hold still. Usually, we use antibodies to find these targets, but they're pricey and don't last long. This new method is a budget-friendly alternative.
They used AI tools, RFdiffusion and ProteinMPNN, to build these proteins. It's like making a custom key for a lock, using only a little info about the lock. This approach is huge because it means we can now spot complex molecules important for health and the environment.
They even tested this in the lab, proving it can find tiny amounts of these targets in human blood. Plus, these proteins can take the heat, literally, which is great for real-world use. They plugged one of these proteins into a sensor, and it made the sensor's signal 21 times stronger when it found its target.
Using AI to diagnose autism in children
A team of Korean pros, including child psychologists and tech experts, have cooked up a smart AI tool that can spot autism in kids. They taught this AI to pick up on unique eye patterns in autistic children, and it nailed diagnosing all 958 kids in the test, half of whom had autism. The AI could also guess how severe the autism was, but it wasn't as sharp there, hitting right about half the time.
This is a big deal because catching autism early means better help for those kids. The researchers are now keen to see if this AI magic works on younger tykes too, since it's only been tested on kids aged 4 to 18 so far.
Nvidia Staffers Warned CEO of Threat AI Would Pose to Minorities
Nvidia employees, specifically Masheika Allgood and Alexander Tsado, raised concerns regarding the potential biases in artificial intelligence (AI) and its impact on minorities. During a meeting with Nvidia CEO Jensen Huang, Allgood and Tsado expressed their fears that AI technology, particularly in areas like facial recognition and self-driving cars, might not adequately recognize or respond to people of color, posing a greater risk to these communities.
Despite their efforts to communicate these issues, Allgood and Tsado felt that their concerns were not taken seriously enough by Huang. They were particularly worried about the lack of diversity at Nvidia, as it could lead to biases in AI development. This concern is supported by data showing Nvidia's low representation of Black and Hispanic employees compared to other tech companies.
Following the meeting, both Allgood and Tsado decided to leave Nvidia, citing the company's failure to prioritize the issues they raised as a key reason. Since their departure, Nvidia has reportedly taken steps to address these concerns, including the appointment of Nikki Pope to lead its Trustworthy AI project and efforts to diversify datasets to reduce bias in AI models.
The broader context of this issue highlights the ongoing challenges in the tech industry, where underrepresentation of minorities can lead to biases in AI systems. This concern is not limited to Nvidia; it reflects a broader industry-wide issue where the lack of diversity can result in AI technologies that do not adequately consider the needs and safety of all communities.
Develop Your First AI Agent: Deep Q-Learning
This article guides you step-by-step to build your own AI agent using Deep Q-Learning. It's aimed at folks with some Python know-how, but you don't need to be an AI whiz. The project? A reinforcement learning gym made from scratch. It’s not just theory; you'll get your hands dirty making an environment, an agent, and setting up the learning process.
First, it breaks down reinforcement learning basics – how agents learn from actions to achieve goals, and introduces Deep Q-Learning. Next, it dives into creating your gym. You'll set up the environment, where your AI will live and learn. The agent's brain, a neural network, comes next, teaching it to make smart moves. Then, it's all about training: your agent learns from experiences, improving over time. You'll also fiddle with learning rates and explore how tweaking them impacts learning.
8 AI TOOLS FOR FREELANCERS RUNNING A SIDE HUSTLE IN 2024
Freelancers juggling side hustles in 2024, listen up! This gig can be a real bear, what with all the extra work beyond your main job. But fear not, AI's got your back with some cool tools:
Lyne AI – Cranks out personalized email openers to get more replies, plus digs up email addresses from LinkedIn.
Logopony – Makes slick logos easy-peasy, no design skills needed.
Scheduler AI – Books meetings like a pro, handling the nitty-gritty details.
Supercreator AI – Your go-to for creating awesome, educational short videos.
ChatGPT – The Swiss Army knife of AI, helping with everything from writing emails to brainstorming ideas.
Receipt Cat – Tracks your spending without the headache, keeping you IRS-ready.
Simplified AI – Cranks out all kinds of content, perfect when you're stuck or not feeling creative.
Canva – Your shortcut to eye-catching social media graphics and presentations.
Bottom line: These AI helpers can save you time and keep you sane in the freelance hustle. Give 'em a whirl!
What'd you think of today's edition? |
What are MOST interested in learning about AI?What stories or resources will be most interesting for you to hear about? |
latest updates on GPT-4, unraveling the complex web of misconceptions and truths in the AI world
Reply