There is a lot of news of artificial intelligence this week from Google, Microsoft, Openai and Anthropic, which we will cover in the news section below. Most of the product innovations that these companies publish are designed at the top of the main “basis” models. These are large models of artificial intelligence that can, once trained, perform all different types of tasks. Big language models are trained today to predict the following word in the sentence, but after this training it can perform many linguistic tasks – from translation to answering questions such as virtual encyclopedias to summary.
But there are still some advantages in training the basic models designed narrowly for specific areas. For example, Google DeepMind’s Alfafold 3 It is the basis for biology. You cannot write poetry. But it can expect the structure of proteins and interactions between two protein or between any protein and any small molecule. This makes it very useful for tasks such as drug design. Wayve, UK self -driving company, Basic models were built You can deal with many different sides of driving – identifying things, identifying the best way to direct the car, and work on the accelerator and brakes, for example. Robotat company that built material intelligence Robotat Foundation models This can help any type of robot perform all different types of tasks without any additional training.
For companies, it is often easier to see a way to the return on investing from these basic models that are somewhat more narrow than it is fully from LLMS. The Swiss army knife is great. But you may not want to use it for surgery. In today’s eye on artificial intelligence, I want to present to you gripSilicon Valley Company, which designed a key model that is supposed to facilitate something sitting at the heart of work decisions: take precise predictions.
Save time, data and money
Usually, the predictions of data require a strenuous work by data scientists, for days, weeks, or even months. Automated learning and deep learning-the most associated machine learning branch has been applied to Amnesty International today-on predictive analyzes for years. But these models are usually designed to create only one type of prediction in one specific context and must be trained in a large group of data for use before you can provide accurate predictions. Major technology companies and main retailers have groups of data needed to train these types of predictive artificial intelligence models. But many smaller companies do not.
The new RFM can be from Kumo – which announces and allows customers today – on the other hand, on the other hand, can handle all types of different predictions. From chewing the customer to credit credit risks to opportunities that the patient who was discharged from the hospital needs to be accepted within 24 hours – kumorfm can deal with all these different predictions, and he can do this almost immediately, without any additional training. “Through the foundation model, it goes to your data, and determines what you mean by Churn, and the second later, you get prediction,” Jure Leskovec, the computer world at Stanford University, who released Komo, told me three years ago, and works as its chief scientist, using an example of creating a customer model. He said that the customer can increase the form of the form on his own data and obtain a 10 % improvement in the accuracy of his predictions.
Using the nerve graphs to discover the main links
The Kumo model depends on Leskovec research in nerve networks, which can encrypt the relationships between things in the network structure, and apply this method to the data listed through different schedules and understand how to change data in these tables over time. (RFM, the name of the Kumo model, symbolizes the Apartol Foundation model.) The model also flourishes the structure of the graph with the same type of transformer structure – which is good in knowing the data that must be paid attention to in order to provide an accurate prediction, even if the decisive data of the rear occur in the back. The foundation model has been trained on the data available to the public as well as what Leskovec said is a large amount of artificial data.
He said that as long as the stamps through the user tables are correct, the Como model is able to provide very accurate predictions. In the standard tests conducted by Kumo, RFM works without any refinement better than some traditional learning methods, equally or better than the human world of data that has photographed a model manually, and slightly worse than a depositary neural network that was specifically trained for this task. By additional adjustment, RFM is performed equally or for some tasks, much better than the nerve network trained in the traditional way of one task. Decally, when compared to trying to use the Meta Llama 3.2b Llama language model and this is required to try to create predictions based on a wave, RFM performance was significantly better. (Standard Komo results have not been independent and verified.)
Kumorfm results can be more interpretation than many models that have been handled by data analysts. This is because human data analysts sometimes develop signs they believe are predicted – for example, saying that the customer may be more likely to buy a specific product if they see an advertisement after 10 pm – but it turns into false. “We can only explain models today through the signals I created. But in our case, we can reach the initial data and say, because of these events, because of this information, we made this decision,” Leskovic said.
Kumo has received $ 37 million of investment capital financing so far from investors including Sequoia Capital, and is currently employing a team of about 50 people divided between Silicon Valley and Europe. Their models have been used so far by companies including the DOORDASH and Reddit Food Series and the Grocery Series in the UK SAINSBRY’s, among other things.
For institutions that are fighting to extract value from their data and frustration due to the long process of building predictive models, Kumo’s approach can represent a major penetration of efficiency. (Amazon Web Services offers a basic model called Chronos to create predictions about things that occur in a time sequence, but it still requires accurate adjustment to achieve accurate results. Datadog data surveillance programs company also provides a similar basic model called TOTO. However, it is the rest of the news of artificial intelligence for this week.
However, here is more news of artificial intelligence.
Jeremy was
[email protected]
Jeremyakahn
Before we get to the news, the latest list of the strongest Fortune Airt is today, and includes a number of important figures for the industrial intelligence industry, including the CEO of AMD Lisa Su, Vice President of Huawei Meng Wanzhou, Anthropor Danieli, Machines Machines and CEO Mira Machi Moratti. You can check the menu here. There is also a great interview with the CEO of New York Times Meredith Kopit Levien by luckRuth Omoh, who affects how the publisher, Amnesty International sees an opportunity and threat, and why Openai. You can check this here.
This story was originally shown on Fortune.com
https://fortune.com/img-assets/wp-content/uploads/2025/05/Edit-Kumo-Headshots-0047-e1747753871770.jpg?resize=1200,600
Source link