OpenAI released a new version of GPT-4 a couple of weeks ago: the 4o model, short for “Omni.” This latest model improves the processing of text, images, and audio, both in input (prompt data) and output (generated content). It also boasts better support for languages other than English, all with reportedly improved performance and halved costs. We asked Sébastien Landeau, CIO of VirtualExpo, what ChatGPT 4o could bring to an e-commerce company.
OpenAI made headlines again last week with the launch of its latest AI model, GPT-4o, derived from the Latin “omnis,” meaning “all.” It is designed to understand and respond in real-time to “all” inputs, including text, audio, and images. Essentially, it can engage in conversations with users, not just provide answers to questions.
Enhanced Performance
Among the key technological advancements of GPT-4o is its total multimodality. The new model can accept and produce text, audio, and image responses.
It also boasts a quick response time. It can react to audio inputs in as little as 232 milliseconds, with an average of 320 milliseconds. This is similar to human response times in conversation.
Additionally, GPT-4o offers improved performance for non-English languages, surpassing GPT-4 Turbo in English text and coding performance while making significant strides in other languages.
GPT-4o also excels in audio and visual understanding compared to existing models. Its vision mode can comment on video streams in real-time. It can even add humorous remarks.
Early user tests have been impressive. For instance, when analyzing the same photo with the same prompt, the new model provides a more detailed understanding of the image than GPT-4. It can for example accurately identify the surface the subject is on and even guess the photo’s location correctly.
Cost Optimization
But what can this new model bring to a B2B company? We asked Sébastien Landeau, CIO of VirtualExpo Group, about the potential positive impacts of such advancements. VirtualExpo is developing several online shops and aims to streamline and scale up the process of adding new products to their e-commerce sites. The company already uses ChatGPT 4.
According to Mr. Landeau, ChatGPT 4o opens numerous opportunities to optimize VirtualExpo shops’ onboarding process cost and performance. He also highlights the strategic need for any company using AI to manage costs.
What do you think of Openai’s new model?
Sébastien Landeau: “This is the culmination of successive improvements to their models. They claim it enhances language processing for non-English languages and doubles performance. What’s really interesting for a company like ours, which uses these tools primarily for product onboarding in our e-commerce business development, is that it cuts costs by half. AI tools are fantastic but we should not forget that they are expensive. For example, OpenAI used to charge us 10$/M tokens for inputs and $30 per million tokens for outputs. For product categorization, we have to include our shop taxonomy in the prompt, which OpenAI also charges for. With 4o, we’ve reduced costs by nearly 50%. They charge us around $5 per million tokens for inputs and $10 per million tokens for outputs. Onboarding a product used to cost us 50 cents; now it’s down to 30 cents per product.”
How do you explain this cost reduction?
Sébastien Landeau: “I think they’re optimizing their model. They’re factoring in their infrastructure and have reduced the number of models. They’re democratizing it. And I think they need to remain competitive to avoid losing customers to competitors like Anthropic or Mistral.”
Is it difficult to keep up with the updates?
Sébastien Landeau: “Evolution is both exciting and challenging. OpenAI is continuously upgrading in terms of performance and programming. I think we’ve had to refactor our code for the third time to keep up with the continuous updates and improvements they are making. This is intrinsic to any new development emerging from research. It’s a booming field, it’s evolving daily, and this is normal. That’s why we created an internal tech watch unit to monitor these changes. After the 4o release, my teams updated the code immediately. OpenAI excels at providing up-to-date documentation on how to use their GPT assistant and knowledge files in agents. Each OpenAI release comes with updated conversational clients, APIs, and excellent documentation.”
And in terms of reliability, what are your impressions so far?
Sébastien Landeau: “These tools are based on large language models (LLMs), so there will always be some level of ‘hallucination’ or unexpected outputs. Achieving perfect industrial processes is an idealistic goal because it’s a statistical model with inherent randomness. We always adopt a QA approach, either manually testing or having one AI analyze another AI’s results. For instance, we use GPT-4 for product categorization and GPT-3.5 Turbo to verify those results. We then conduct manual acceptance tests, resulting in fairly reliable outcomes.”
Is it strategic to rely solely on OpenAI?
Sébastien Landeau: “Putting all your eggs in one basket is complicated. If there’s maintenance or a service outage, how do we continue our processes? One possibility is to mix SaaS technologies, like Anthropic or Mistral. Another strategy to reduce dependency is integrating neural networks and training them on our data. The advantage is the low per-unit cost compared to online services. We’re experimenting with neural networks for image analysis and training models with our shop images. While not perfect yet, the results are promising. Integrating internal LLMs, especially open-source models like Mistral, is in our backlog. It requires significant computational resources for training, but this is necessary to manage costs. Using OpenAI for categorization, acceptance, image vision, and attribute evaluation leads to multiple calls per product, quickly becoming expensive. This could become costly and difficult to manage for onboarding millions of products. So, strategically, we must keep costs manageable and therefore use OpenAI for limited volumes.”
Would you say AI enables time optimization or cost optimization?
Sébastien Landeau: “So far, we haven’t achieved budget optimization using AI; we have only optimized time. The target ratio between manual processing and AI is 20 to 80 in favor of AI. But, with this new GPT-4o model, we are definitely seeing a cost reduction. Besides, from a technical standpoint, a developer being stuck on technical problems for three days no longer happens today thanks to these tools.”
![Image [BUYING GUIDE] How to Choose the Right Shelving?](/wp-content/uploads/sites/3/Featured-7-320x213.jpg)
