That is why Marhosini is using AI to improve AI chips. Back in 2021, he and his colleagues in Google made a Non -LLMAI System It can decide where to improve performance computer computer chips to put different ingredients on the chip. Although some other researchers Unable to duplicate. The results of this study, Mirozini say Nature This article was investigated and maintained the authenticity of work – and it notes that Google has used system design for several generations of custom AI chips.
Recently, Marozini has applied the LLM on the issue of writing, lower -level functions, which control that various operations, such as matrix multiplication, are conducted in chips. It has found that even the LLM of the general purpose, in some cases, can write Dana that Racing fast More than human designed version.
Elsewhere in Google, scientists created a system that they used to improve the company’s LLM infrastructure. This system, called Alphalololol, indicates Google’s Gemini LLM to write algorithms to solve some problems, reviews these algorithms, and calls for Gemini to improve the most successful – and it is implemented several times. Alfolo developed a new approach to running data centers, which saved Google’s computational resources 0.7 %, further improved Google’s customs chip design, and designed a new kernel that increased Gemini training by 1 %.
It may feel like a small improvement, but in a huge company like Google, it is equivalent to a lot of time, money and energy saving. And Google Deep Mind’s staff research scientist Matage Blog led the Alfaulovolo project, saying he and his team examined the system on just a small component of the Gemini training pipeline. It says implementing it more widely can save further.
3. Automatic training
LLMS famous data are hungry, and their training is expensive at every stage. In certain domains-non-minor programming languages, for example-the real world data is little to effectively train LLM. Learning reinforcements with human feedback, a technique in which humans face LLM reactions to the indicator and then trained LLM using these scores, they have been the key to creating models that treat human standards and preferences, but are slow and human influences.
Fast, LLM is being used to fill the space. If many examples are indicated, LLMS can produce comprehensible artificial data in domains that they have not been trained, and then artificial data can be used for training. LLMS can also be used effectively to learn reinforcement: In a view called “LLM Judge LLM” instead of humans, they are used to score the output of models that are being trained. This approach is the key to the influential “Constitution AI” framework proposed by anthropic researchers in 2022, in which one LLM is trained to be less harmful based on the opinion of another LLM.
Data deficiencies for AI agents is a particularly serious problem. Effective agents need to be able to perform multi -stop projects to accomplish special tasks, but examples of the successful completion of step -by -step task are low online, and it would be expensive to use humans to prepare new examples. To overcome this limit, Stanford’s Marozini and his colleagues have recently done a pilot Technique In which an LLM agent develops a possible step -by -step approach to a given problem, an LLM judge estimates whether every step is correct or not, and then a new LLM agent is trained at these steps. “You are no longer limited by statistics, because the model can only create more and more experiences,” says Merhosini.
4. Full agent design
An area where the LLM has not yet made major partnerships is in the design of LLM itself. Today’s LLM are all based on a nervous network structure called a transformer, which was proposed by human researchers in 2017, and subsequently a significant improvement in architecture was also human designs.