Embam: 5 methods that institutions can reduce the costs of artificial intelligence without sacrificing performance

1: The right size of the mission model

Avoid undermining giant models and general purposes for each use. The models of the mission or the cutting can match, or even Transcend, Larger models In terms of accuracy of the targeted work burdens – at a lower cost and with low energy consumption.

In fact, Luccioni is found in a test that the task model uses energy less than 20 to 30 times of general purpose energy. She said: “Since it is a model that can do this task, unlike any task you throw, which these are often large linguistic models,” she said.

The distillation is the key here; The full form can be trained at first from scratch and then polish it for a specific task. “Deepseek R1, for example,” is so huge that most organizations cannot use them “because you need at least 8 official processing units. On the contrary, distilled versions can be 10, 20, or even 30x smaller and run on a single graphics processing unit.

She pointed out that open source models help in efficiency, because they do not need training from scratch. This is compared to only a few years, when companies were waste resources because they were unable to find the model they needed; Nowadays, they can start with a basic model, adjust and adapt it.

“It provides a gradual joint innovation, unlike that everyone is trained in their models on their data collections and mainly wasting the account in this process,” said Luxyoni.

It has become clear that companies are very disappointed General AIBecause the costs are not yet proportional to the benefits. Public use cases, such as writing emails or copying meeting notes, really useful. However, the task models still require “a lot of work” because the models outside the box do not cut them and are also more expensive, as Luccioni said.

These are the following limits of the added value. “Many companies want to do a specific task.” “They do not want AGI, they want a specific intelligence. This is the gap that must be fed.”

2. Make the default efficiency

Adoption of “payment theory” in system design, conservative thinking budgets, always reduce obstetric features and require adherence to high -cost calculation patterns.

In cognitive sciences, “defense theory” is the approach to managing behavior changes designed to influence human behavior with skill. Lucioni pointed out that the “ecclesiastical example” adds table tools to eating abroad: to make people decide whether they want plastic utensils, instead of automatically including them with each request, can significantly reduce waste.

“Just to make people choose something for cancellation of something in fact, it is actually a very strong mechanism for changing people’s behavior,” said Luxyoni.

Virtual mechanisms are also unnecessary, as they increase use, and therefore, costs because models do more than they need. For example, with common search engines like Google, the Gen Ai summary is automatically filled at the top. Lucioni also noticed that when I recently used the GPT-5 of Openai, the model automatically works in a full thinking mode on “very simple questions”.

“For me, the exception should be,” she said. Like, “What is the meaning of life, then for sure, I want a summary of Gen Ai. But with “What is the weather in Montreal” or “What are the working hours in the local pharmacy?” I do not need a summary of artificial intelligence, however it is default.

3. Improving the use of devices

Use the cylinder. Adjust the exact payments sizes and ensure them to generate the specific devices to reduce the wasted memory and withdraw the energy.

For example, the companies themselves should ask: Should the model be all the time? Will people assemble it in the actual time, 100 requests at one time? In this case, improvement is always necessary, as Lucioni indicated. However, in many others, it is not; The model can be operated periodically to improve memory use, and misinformation can ensure optimal memory use.

“It is somewhat similar to the engineering challenge, but it is a very specific challenge, so it is difficult to say,” just distillation of all models, “or” changing accuracy in all models. ”

In one of its recent studies, I found that the size of the payments depends on the devices, even to the specified type or version. The transition from one batch size to one plus can increase the use of energy because the models need more memory bars.

“This is something that people do not really look, they are like,” Oh, I will enhance the size of the batch, “but it really comes to switching all these different things, and suddenly it is very effective, but it only works in your specified context.”

4. Motivating energy transparency

It always helps when people are motivated; To this end, the embrace was launched earlier this year Artificial intelligence power points. It is a new way to enhance more energy efficiency, using a 1 to 5 -star rating system, with the most efficient models that get a “five -star” condition.

It can be considered “Energy Star for Artificial Intelligence”, and was inspired by the federal program that is likely to end, which sets the specifications of energy efficiency and eligible brands with the energy star logo.

“Two decades ago, this was a really positive motivation. People wanted to classify the stars, right?” Luxiouni said. “Something similar to the power degree will be great.”

Her embedded face The leaders are now,, Which you plan to update with new models (Deepseek, GPT -SS) in September, and do so constantly every 6 months or sooner with the availability of new models. Lukoni said that the goal is that models builders will consider the classification “honor badge.”

5. Reflection on the mentality of “more account is better”

Instead of chasing the largest GPU, start by asking: “What is the smartest way to achieve the result?” For many work burdens, the most intelligent structures and coordinated data outperform the brute force.

“I think people may not need many graphics processing units as they think,” said Luxyoni. Instead of just going to the largest groups, companies urged to rethink the tasks that graphics processing units will complete graphics units and why they need them, how they have made these types of tasks before, and what is the addition of additional graphics processing units in the end.

She said: “It is a kind of this race to the bottom where we need a larger group.” “She is thinking about what she uses for artificial intelligence, and what technology you need, what does that require it?”

Daily visions about business use cases with VB daily

If you want to persuade your boss at work, you have covered VB Daily. We give you the internal journalistic precedence over what companies do with obstetric artificial intelligence, from organizational transformations to practical publishing operations, so that you can share visions of the maximum return on investment.

Read with us privacy policy

Thanks for subscribing. Check more VB news bulletins here.

An error occurred.