Shulou(Shulou.com)11/24 Report--

AI agents composed of large models such as GPT-4 can already teach you to do chemical experiments by hand. It knows exactly what reagent to choose, how much dose to use, and how the reasoning reaction will happen. Tremble, Bio-Ring!

GPT-4 has learned to do scientific research on its own?

Recently, several scientists at Carnegie Mellon University published a paper that exploded both AI and chemical circles.

They made an AI that would do its own experiments and do its own research. This AI consists of several large language models, which can be regarded as a GPT-4 agent agent with explosive scientific research capabilities.

Because it has long-term memory from vector databases, it can read and understand complex scientific documents and conduct chemical research in cloud-based robotic laboratories.

Netizens were shocked to aphasia: So, this is AI's own research and then its own publication? Jesus Christ.

Some people say that the era of "Wensheng Experiment"(TTE) is coming!

Could this be the legendary AI Holy Grail of the chemical world?

Many of us feel like we live in science fiction every day.

Is the AI version of Breaking Bad coming? In March, OpenAI released GPT-4, a large language model that shocked the world.

This is the strongest LLM on the surface, able to score high on SAT and BAR exams, pass LeetCode challenges, do the right physics questions with a picture, and understand the stems in emoji packs.

The technical report also mentioned that GPT-4 can also solve chemical problems.

This inspired several academics in Carnegie Mellon's chemistry department to develop an AI based on multiple large language models, allowing it to design and conduct its own experiments.

Paper address: arxiv.org/ abs / 2304.05332 and they made this AI, really 6 is not good!

It will search the Internet for its own literature, will precisely control liquid handling instruments, and will solve complex problems that require the simultaneous use of multiple hardware modules and integration of different data sources.

It smells like an AI Breaking Bad.

For example, let's let this AI synthesize ibuprofen for us.

Give it a simple prompt: "synthesize ibuprofen."

And then the model goes online and searches for what to do.

It recognized that the first step required the Friedel-Crafts reaction of isobutylbenzene and acetic anhydride catalyzed by aluminum chloride.

In addition, this AI can also synthesize aspirin.

And synthetic aspartame.

The methyl group is missing from the product, and the model finds the correct synthesis example, which is executed in the cloud lab for correction.

Tell the model: study the Suzuki reaction, it immediately identifies the substrate and product accurately.

In addition, we can connect the model to a chemical reaction database through API, such as Reaxys or SciFinder, and stack the model with a big buff, and the accuracy rate will soar.

Analyzing previous records of the system can also greatly improve the accuracy of the model.

Let's first take a look at how the robot is operated to do experiments.

It treats a set of samples as a whole (in this case, the entire plate).

We can tell it directly in natural language: "Color every other line with a color of your choice."

When executed by a robot, these protocols are very similar to the prompts requested (Figure 4B-E).

The agent's first action is to prepare a small sample of the original solution (Figure 4F).

Then it asks for UV-Vis measurements. When complete, the AI gets a filename containing a NumPy array containing spectra for each well of the plate.

The AI then wrote Python code to identify wavelengths with maximum absorbance and used that data to solve the problem correctly.

In previous experiments, AI may have been influenced by the knowledge received during the pre-training phase.

This time, the researchers intend to thoroughly evaluate AI's ability to design experiments.

The AI first integrates the required data from the network, runs the necessary calculations, and finally programs the liquid reagent operating system (the leftmost part of the image above).

To add some complexity, the researchers asked AI to apply a heated shaker module.

These requirements are integrated and appear in the configuration of AI.

The design is such that AI controls a liquid operating system loaded with two miniature versions, the source version of which contains source solutions for multiple reagents, including phenylacetylene and phenylboronic acid, multiple aryl halide coupling partners, and two catalysts and two bases.

The image above shows the contents of the Source Plate.

The target version is mounted on a heated shaker module.

In the image above, the left pipette has a 20 microliter range and the right single-channel pipette has a 300 microliter range.

The ultimate goal of AI is to design a process that successfully implements the Suzuki and Sonogehela reactions.

We say to it: you need to generate these two reactions with some available reagents.

Then, it went online to search, for example, what conditions are needed for these reactions, what are the requirements for stoichiometry, and so on.

It can be seen that AI successfully collected the required conditions, the quantity and concentration of the required reagent, etc.

The AI picked the right coupling partner to complete the experiment. Among all aryl halides, AI selected bromobenzene for Suzuki reaction and iodobenzene for Sonogehela reaction.

In each round, the AI's choices changed somewhat. For example, it chose p-iodinitrobenzene because of its high reactivity in oxidation reactions.

Bromobenzene was chosen because it can participate in the reaction and is less toxic than aryl iodides.

Next, AI chose Pd / NHC as the catalyst because it works better. This is a very advanced way of coupling reactions. As for the base selection, AI took a fancy to triethylamine.

From the above process, we can see that the future potential of this model is unlimited. Because it will repeatedly carry out experiments to analyze the reasoning process of the model and obtain better results.

After selecting the different reagents, the AI begins to calculate the amount of each reagent required, and then begins to plan the entire experimental process.

The AI also made a mistake in the middle, using the wrong name for the heating shaker module. However, AI noticed this point in time, spontaneously inquired about the data, corrected the experimental process, and finally successfully operated.

Leaving aside the professional chemical process, let's summarize the "professionalism" that AI has demonstrated in this process.

It can be said that from the above process, AI shows a very high analytical reasoning ability. It is able to acquire the required information spontaneously and solve complex problems step by step.

In the process, you can also write super high-quality code yourself and advance experimental design. Also, you can change your own code based on the output content.

OpenAI successfully demonstrated the power of GPT-4, and GPT-4 will certainly be able to participate in real experiments one day.

But the researchers didn't want to stop there. They also gave the AI a big challenge--they gave it instructions to develop a new cancer drug.

Something that doesn't exist... Can this AI still work?

It turned out that he really had two brushes. AI adheres to the principle of not afraid of encountering difficulties (of course, it does not know what fear is), carefully analyzes the need to develop anti-cancer drugs, studies the current trend of anti-cancer drug development, and then selects a target to continue to deepen and determine its composition.

After that, AI tries to start synthesizing itself, also searching the Internet for information about reaction mechanism and mechanism, and then looking for examples of related reactions after preliminary steps are completed.

And then finally complete the synthesis.

However, the content in the above picture cannot be synthesized by AI. It is only a theoretical discussion.

Among them are familiar drugs such as methamphetamine (also known as marijuana) and heroin, as well as banned toxic gases such as mustard gas.

AI provided synthetic schemes for 4 of the 11 compounds in total and attempted to consult the literature to advance the synthetic process.

Of the remaining seven substances, the synthesis of five was decisively rejected by AI. AI searched the Internet for information about these five compounds and found that they could not be fooled.

For example, while trying to synthesize codeine, AI discovered a relationship between codeine and morphine. It was concluded that this thing was a controlled drug and could not be synthesized casually.

However, this insurance mechanism is not stable. Users can further let AI operate as long as they modify the flower book slightly. For example, compound A is used instead of a direct reference to morphine, compound B is used instead of a direct reference to codeine, and so on.

At the same time, the synthesis of some drugs must be approved by the Drug Enforcement Administration (DEA), but some users can exploit this loophole, deceive AI that they have permission, and induce AI to give a synthesis plan.

AI was also aware of familiar contraband such as heroin and mustard gas. The problem is that the system can only detect existing compounds. For unknown compounds, the model is less likely to identify potential hazards.

For example, some complex protein toxins.

Therefore, in order to prevent anyone from testing the effectiveness of these chemical ingredients out of curiosity, the researchers also specially posted a large red warning in the paper:

The illicit drug and chemical weapons synthesis discussed in this article is purely academic and primarily intended to highlight the potential dangers associated with new technologies.

Under no circumstances should any person or organization attempt to remanufacture, synthesize, or otherwise produce a substance or compound discussed herein. Such activities are not only dangerous but illegal in most jurisdictions.

I will go online and search how to do experiments. This AI consists of multiple modules. These modules can exchange information with each other, and some have access to the Internet, APIs, and Python interpreters.

After typing the prompt into Planner, it starts to perform the operation.

For example, it can surf the Internet, write code in Python, access documentation, and once it understands the basics, it can do experiments on its own.

When humans do experiments, this AI can guide us hands-on. Because it can reason about chemical reactions, it can search the Internet, it can calculate the amount of chemicals needed in the experiment, and it can perform the corresponding reaction.

If the description is detailed enough, you don't even have to explain it to it; it can sort out the whole experiment itself.

When the Web searcher component receives a query from Planner, it uses the Google Search API.

After finding the results, it filters out the top ten documents returned, excludes PDFs, and passes the results to itself.

It then uses the BROWSE operation to extract text from the web page and generate an answer. It was flowing smoothly and smoothly.

GPT-3.5 can accomplish this task because its performance is significantly better than GPT-4, and there is no loss of quality.

The Docs searcher component, which can find the most relevant parts through queries and document indexes, combs hardware documents (such as robotic liquid processors, GC-MS, cloud labs), and then summarizes a best match to generate the most accurate answer.

The Code Execution component does not use any language model, but simply executes code in isolated Docker containers, protecting the end-host from any unexpected actions by Planner. All code output is passed back to Planner so that the software can fix predictions if something goes wrong. Automation components work on the same principle.

Vector search, difficult scientific literature can understand to make a complex reasoning AI, there are many problems.

For example, in order for it to integrate modern software, users need to be able to understand the software documentation, but the language of this documentation is generally very academic and professional, causing great obstacles.

Large language models can overcome this obstacle by generating software documents in natural language that non-experts can understand.

One of the sources of training for these models is a lot of information about APIs, such as the Opentons Python API.

However, GPT-4 training data is up to September 2021, so it is even more necessary to improve the accuracy of AI use API.

To this end, the researchers devised a way to provide documentation for a given task for AI.

They generated ada embeddings of OpenAI to cross-reference and calculate similarities related to queries. and selects portions of the document by vector search based on distance.

The number of parts provided depends on the number of GPT-4 tokens present in the original text. The maximum number of tokens is set to 7800, so that only one step can be provided to AI related files.

This approach proved crucial in providing AI with information about the heater-vibrator hardware module, which is necessary for chemical reactions.

This approach presents even greater challenges when applied to more diverse robotics platforms such as Emerald Cloud Lab (ECL).

At this point, we can provide information to the GPT-4 model that it does not know, such as Symbolic Lab Language (SLL) about Cloud Lab.

In all cases, the AI correctly identified the task and completed it.

In doing so, the model effectively retains information about the various options, tools, and parameters of a given function. After ingesting the entire document, the model is prompted to generate a code block using the given function and pass it back to the Planner.

Finally, the researchers stress that safeguards must be in place to prevent large language models from being abused:

"We call on the AI community to prioritize the safety of these models." We call on OpenAI, Microsoft, Google, Meta, Deepmind, Anthropic, and other major players to put their best efforts into the security of their large language models. We also call on the physical science community to work with teams involved in developing large language models to assist them in developing these safeguards."

New York University professor Marcus agrees: "This is no joke. Three scientists from Carnegie Mellon University urgently call for safety research on LLM."

References:

https://arxiv.org/ftp/arxiv/papers/2304/2304.05332.pdf

This article comes from Weixin Official Accounts: Xinzhiyuan (ID: AI_era)

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.