Suggestions
Igor Akimov
AI Solutions Product Manager in Nicosia, Cyprus
Igor Akimov is an accomplished AI Solutions Product Lead at Wrike, where he has been instrumental in leading cross-functional teams to deliver AI-driven solutions, including advancements in search, machine learning (ML), natural language processing (NLP), and generative AI. His tenure at Wrike spanned from April 2020 to December 2023, during which he championed the rollout of the "Work Intelligence" suite, integrating AI-enhanced features that improved task management and risk prediction capabilities within the platform. Akimov is dedicated to democratizing AI, aiming to make its benefits accessible to all teams, not just experts.1
In addition to his role at Wrike, Akimov has taken on multiple positions as an AI Product Owner and trainer, focusing on enhancing product management efficiency through AI. His previous experience includes directing product strategies at ABBYY, where he managed mobile applications and led innovative projects that significantly increased the team's size and profitability.1
Overall, Akimov combines a passion for technology with a commitment to fostering innovation and efficiency in product management.
Highlights
Very cool news from OpenAI! They gave the ability to strictly control the model's response, which is critically important for all sorts of multi-agents and functions.
Previously, even with the "produce JSON" checkbox, the responses might not match the schema you requested, which then resulted in additional checks, re-requests, backend rework, and so on. Now you can enable "strict: true" in the function calling on any model and it will only give a response in the format you need.
And if you use the latest gpt-4o-2024-08-06 model or the mini version, you can drop json_schema into your request, enable "strict: true" and the response will 100% (at least according to their tests) match your data schema, that is, all the necessary fields will be filled in the correct order and with the correct types. This, of course, really simplifies development.
A new property, "refusal", has appeared, which is passed if the model refuses to respond (again, previously it had to be detected on the backend). They updated their SDKs on Python and Node to support all this.
Use-cases they describe:
- creating UI dynamically, according to your templates
- separating reasoning and the answer itself (they gave an example with a comparison of 9.9 and 9.11, haha)
- extracting structured data from unstructured
They achieved amazing results in 2 ways: fine-tuned LLM, and then polished with some extra checks, quite interesting ones: they write a set of language rules for each scheme - context-free grammar (CFG), like: At the beginning of JSON there should be "{", then a line break and the word "type", then one of the types, a line break, a value from such and such limits, and so on. As a result, after each token, the model is shown what the next ones can be, so that it does not try to hallucinate.
Sounds really technological and cool. As a result, according to their tests, they achieved a 100% valid answer all the times.
The only thing is that the first request with a new scheme can take from 10 seconds and even closer to a minute, but then they cache the scheme and subsequent ones with the same scheme go quickly.
More details here: https://lnkd.in/dr3sq_Re
And for dessert, the new model gpt-4o-2024-08-06 costs 2 times less for input and a third less for output. Instead of $5/$15 - $2.5/$10. Wow!
Great test of LLM's suitability for RAG. The results are quite surprising.
In first place for short context (up to 5000 tokens, a few pieces from the knowledge base) is Claude Sonnet 3.5, Opensource - Meta LLama-3-70b Instruct.
For medium context (5000-25000 tokens, 20 chunks or the whole document) - Gemini-1.5 Flash (wow), from Opensource - qwen2-72b-instruct
For a context larger than 40000 tokens (several documents in full) - Claude 3.5 Sonnet. There is no good Opensource solution with such a large context.
Looks like need to switch from GPT-4o to Sonnet 3.5 or Gemini 1.5
https://lnkd.in/gcM4vgcT
