Page 158 - CW E-Magazine (31-10-2023)
P. 158

News from Abroad


       ROOM FOR IMPROVEMENT

       Japanese scientists say GPT-4 shows promise in aiding

       chemistry research, but limitations remain

          GPT-4, the latest version of the arti-                          cally been  trained  on.  Overall, it was
       ficial  intelligence  (AI)  system  from                           able to outperform some existing com-
       OpenAI, the developers of Chat-GPT,                                putational algorithms, but fell short
       demonstrates considerable usefulness                               against others.
       in tackling chemistry challenges,  but
       still has significant weaknesses.                                     “The results indicate that GPT-4 can
                                                                          tackle a wide range of tasks in chemical
          GPT-4, which stands for Generative                              research, spanning from textbook-level
       Pre-trained  Transformer 4, belongs to                             knowledge to addressing untrained
       a category of  large language models                               problems and optimising multiple vari-
       within  the  field  of  AI.  These  models  capabilities, the researchers set the sys-  ables,” said Mr. Hatakeyama-Sato. “In-
       can gather and analyse vast quantities  tem a series of chemical tasks focused  evitably, its performance relies heavily
       of  information  to  solve  user-defined  on organic chemistry.  These covered  on the quality and quantity of its train-
       problems. A notable advancement with  basic chemical theory,  the handling  ing data, and there is much room for
       GPT-4 is its capability  to utilise  both  of molecular data, predicting the pro-  improvement in its inference capabili-
       text and image data for analysis.  perties of chemicals,  the outcome  of  ties,” he noted.
                                         chemical processes and proposing new
          “It has a notable understanding of  chemical procedures.           The researchers  emphasised that
       chemistry, suggesting it can predict and                           their work was only a preliminary in-
       propose experimental results in ways   The results of the investigation  vestigation, and  that future  research
       akin to human thought processes,” said  were varied,  revealing  both strengths  should broaden the scope of the trials
       chemist,  Mr. Kan Hatakeyama-Sato,  and significant limitations. GPT-4 dis-  and dig deeper into the performance of
       at  the Tokyo  Institute  of Technology.   played a good understanding of general  GPT-4 in more diverse research  sce-
       Mr. Hatakeyama-Sato and his colleagues  textbook-level  knowledge in organic   narios. They also hope to develop their
       have discussed the potential of GPT-4   chemistry. It was weak, however, when  own large language models specialising
       in chemical  research in the journal   set tasks dealing with specialised con-  in chemistry and explore their integra-
       Science and  Technology of  Advanced  tent or unique methods for making   tion with existing techniques.
       Materials: Methods.               specific organic compounds. It displayed
                                         only  partial  efficiency  in  interpreting   “In the meantime, researchers
          Although the specific datasets used  chemical  structures  and converting  should certainly consider applying
       for training GPT-4 have not been dis-  them into a standard notation. One in-  GPT-4 to chemical challenges, possibly
       closed by its developers, it has clearly  teresting feat was  its  ability to  make  using hybrid methods that include
       learned a significant amount of detailed  accurate predictions for the properties  existing specialised techniques,” said
       chemistry  knowledge.  To analyse its  of  compounds  that  it  had  not  specifi-  Mr. Hatakeyama-Sato.

       REWARDING INNOVATION
       In-cosmetics Asia 2023 awards finalists announced

          In-cosmetics  Asia  has announced   The  20  shortlisted ingredients se-  the greatest number of entries, closely
       the shortlist for this year’s ‘Innovation  lected  by a panel  of judges, will  vie  followed by haircare and suncare.
       Zone’ awards celebrating cosmetic sci-  for the ‘Innovation  Zone  Best Active  Skincare  submissions featured  a wide
       ence excellence across the Asia-Pacific  Ingredient’ and ‘Innovation Zone Best  range of active  ingredients and green
       region.  The  leading cosmetic  ingredi-  Functional Ingredient’ awards.  solutions that cater to the diverse and
       ents trade fair will be held in Bangkok                            evolving needs of  consumers across
       from November 7-9.                  Skincare products top the list for  the region. Qualifying ingredients


       158                                                                   Chemical Weekly  October 31, 2023


                                      Contents    Index to Advertisers    Index to Products Advertised
   153   154   155   156   157   158   159   160   161   162   163