Liquid neural networks are inspired by biological neurons to implement algorithms that remain adaptable even after training. [Hasani] demonstrates a machine vision system that steers a car to perform lane keeping with the use of a liquid neural network. The system performs quite well using only 19 neurons, which is profoundly fewer than the typically large model intelligence systems we’ve come to expect. Furthermore, an attention map helps us visualize that the system seems to attend to particular aspects of the visual field quite similar to a human driver’s behavior.
[Mathias Lechner] and [Ramin Hasani]
The typical scaling law of neural networks suggests that accuracy is improved with larger models, which is to say, more neurons. Liquid neural networks may break this law to show that scale is not the whole story. A smaller model can be computed more efficiently. Also, a compact model can improve accountability since decision activity is more readily located within the network. Surprisingly though, liquid neural network performance can also improve generalization, robustness, and fairness.
A liquid neural network can implement synaptic weights using nonlinear probabilities instead of simple scalar values. The synaptic connections and response times can adapt based on sensory inputs to more flexibly react to perturbations in the natural environment.
We should probably expect to see the operational gap between biological neural networks and artificial neural networks continue to close and blur. We’ve previously presented on wetware examples of building neural networks with actual neurons and ever advancing brain-computer interfaces.
In this article, we will compare four of the most advanced AI bots: GPT-4, Bing, Claude+, Bard, and GitHub Co-Pilot. We will examine how they work, their strengths and weaknesses, and how they compare to each other.
Testing the AI Bots for Coding
Before we dive into comparing these four AI bots, it’s essential to understand what an AI bot for coding is and how it works. An AI bot for coding is an artificial intelligence program that can automatically generate code for a specific task. These bots use natural language processing and machine learning algorithms to analyze human-written code and generate new code based on that analysis.
To start off we are going to test the AI on a hard Leetcode question, after all, we want to be able to solve complex coding problems. We also wanted to test it on a less well-known question. For our experiment, we will be testing Leetcode 214. Shortest Palindrome.
[…]
GPT-4 is highly versatile in generating code for various programming languages and applications. Some of the caveats are that it takes much longer to get a response. API usage is also a lot more expensive and costs could ramp up quickly. Overall it got the answer right and passed the test.
[…]
[Bing] The submission passed all the tests. It beat 47% of submissions on runtime and 37% on memory. This code looks a lot simpler than what GPT-4 generated. It beat GPT-4 on memory and it used less code! Bing seems to have the most efficient code so far, however, it gave a very short explanation of how it solved it. Nonetheless, best so far.
[…]
[Claude+] The code does not pass the submission test. Only 1/121 of the test passed. Ouch! This one seemed promising but it looks like Claude is not that well suited for programming.
[…]
[Bard] So to start off I had to manually insert the “self” arg in the function since Bard didn’t include it. From the result of the test, Bard’s code did not pass the submission test. Passing only 2/121 test cases. An unfortunate result, but it’s safe to say for now Bard isn’t much of a coding expert.
[…]
[Github CodePilot] This passes all the tests. It scored better than 30% of submissions on runtime and 37% on memory.
It’s fun, you can see the coding examples (with and without comments) that were output by each AI in the link
Anyone can use ChatGPT for free, but if you want to use GPT4, the latest language model, you have to either pay for ChatGPT Plus, pay for access to OpenAI’s API, or find another site that has incorporated GPT4 into its own free chatbot. There are sites that use OpenAI such as Forefront (opens in new tab) and You.com (opens in new tab), but what if you want to make your own bot and don’t want to pay for the API?
A GitHub project called GPT4free (opens in new tab) allows you to get free access to the GPT4 and GPT3.5 models by funneling those queries through sites like You.com (opens in new tab), Quora (opens in new tab) and CoCalc (opens in new tab) and giving you back the answers. The project is GitHub’s most popular new repo, getting 14,000 stars this week.
Now, according to Xtekky, the European computer science student who runs the repo, OpenAI has sent a letter demanding that he take the whole thing down within five days or face a lawsuit.
I interviewed Xtekky via Telegram, and he said he doesn’t think OpenAI should be targeting him since he isn’t connecting directly to the company’s API, but is instead getting data from other sites that are paying for their own API licenses. If the owners of those sites have a problem with his scripts querying them, they should approach him directly, he posited.
[…]
On the backend, GPT4Free is visiting various API urls that sites like You.com, an AI-powered search engine that employs OpenAI’s GPT3.5 model for its answers, use for their own queries. For example, the main GPT4Free script hits the URL https://you.com/api/streamingSearch, feeds it various parameters, and then takes the JSON it returns and formats it. The GPT4Free repo also has scripts that grab data from other sites such as Quora, Forefront, and TheB. Any enterprising developer could use these simple scripts to make their own bot.
“One could achieve the same [thing by] just opening tabs of the sites. I can open tabs of Phind, You, etc. on my browser and spam requests,” Xtekky said. “My repo just does it in a simpler way.”
All of the sites GPT4Free draws from are paying OpenAI fees in order to use its large language models. So when you use the scripts, those sites end up footing the bill for your queries, without you ever visiting them. If those sites are relying on ad revenue from their sites to offset these API costs, they are losing money because of these queries.
Xtekky said that he is more than happy to take down scripts that use individual sites’ APIs upon request from the owners of those sites. He said that he has already taken down scripts that use phind.com, ora.sh and writesonic.com.
Perhaps more importantly, Xtekky noted, any of these sites could block external uses of their internal APIs with common security measures. One of many methods that sites like You.com could use is to block API traffic from any IPs that are not their own.
Xtekky said that he has advised all the sites that wrote to him that they should secure their APIs, but none of them has done so. So, even if he takes the scripts down from his repo, any other developer could do the same thing.
[…]
Xtekky initially told me that he hadn’t decided whether to take the repo down or not. However, several hours after this story first published, we chatted again and he told me that he plans to keep the repo up and to tell OpenAI that, if they want it taken down, they should file a formal request with GitHub instead of with him.
“I believe they contacted me before to pressurize me into deleting the repo myself,” he said. “But the right way should be an actual official DMCA, through GitHub.”
Even if the original repo is taken down, there’s a great chance that the code — and this method of accessing GPT4 and GPT3.5 — will be published elsewhere by members of the community. Even if GPT4Free had never existed anyone can find ways to use these sites’ APIs if they continue to be unsecured.
“Users are sharing and hosting this project everywhere,” he said. “Deletion of my repo will be insignificant.”
