“American tech companies dominate the market for AI models
on which applications like ChatGPT are based. Now the German start-up Aleph
Alpha has examined the performance of its own model - with encouraging results.
The ChatGPT software from the American company Open AI has
been causing a sensation for almost three months. ChatGPT can use artificial
intelligence (AI) to answer a wide variety of queries in a previously unknown to it. The tech companies Google and Microsoft, which is involved in Open AI,
are now outdoing each other with new AI announcements almost every week. The
focus is on the United States. 73 percent of the large AI models on which
applications like ChatGPT are based come from America.
But with the start-up Aleph Alpha, there is at least one
promising company in Germany that has already developed its own model called
Luminous. The Heidelberg-based company has now comprehensively tested the
performance of its AI for the first time and came to the conclusion that its
own model can keep up with those of OpenAI, Meta and the open, international
research model Bloom.
The company tested the models using a neutral test procedure
from the open research and development community "Eleuther AI" with
regard to the correct classification, evaluation and creation of texts and the
answering of questions about specific text content. The procedure is also used
by the other major providers to evaluate their software. According to the
evaluation, the accuracy of the currently best model from Heidelberg is an
average of 53.4 percent in the so-called core functions. For comparison: Open
AI’s “davinci” model comes in at 53.7 percent, Meta’s OPT at 52.2 percent and
Bloom at 49.3 percent.
"Important step towards Europe's technological
sovereignty"
"Luminous is a strong alternative in many environments
and thus an important step towards Europe's technological sovereignty,"
said the founder of Aleph Alpha, Jonas Andrulis.
The start-up also emphasizes
that Luminous achieves the results with significantly fewer parameters than the
competition. Parameters are the values that the AI learns during its training
and refines through further training. In fact, the Aleph Alpha model contains
almost 70 billion such parameters. The compared competitor models all work with
almost 175 billion parameters. According to the company, this enables Aleph
Alpha to use less resources.
This can also be an economic advantage. Because the
operation of large AI models requires a lot of computing power. Experts assume
that each request to ChatGPT therefore costs around 5 cents.
This is one of the
reasons why Aleph Alpha does not offer a mass product for end customers, but
has specialized in applications for companies and public administration.
The
company has built its own AI data center in Bayreuth for this purpose.
In
October 2022, for example, the “Lumi” citizen assistant, which is based on
Luminous, was launched on behalf of the city of Heidelberg. The system is
capable of responding to individual, non-preprogrammed questions from
citizens.”
Komentarų nėra:
Rašyti komentarą