Suggestions
Tomas Vykruta
Founder & President @ EvolutionIQ. Twitter @tvykruta. Helping Injured & Disabled Individuals Return to Work Sooner. Google AI.
Tomas Vykruta is the Co-founder and President of EvolutionIQ, an AI-powered claims guidance platform for insurance companies.14 He has a background in technology and artificial intelligence, having spent nearly a decade working on machine learning and AI projects at Google before co-founding EvolutionIQ in 2019.25
Vykruta's expertise lies in data science and AI applications. He played a crucial role in developing EvolutionIQ's AI-driven claims guidance system, which helps insurers handle claims more effectively and accurately.1 The company's technology has been adopted by 70% of the top 15 U.S. disability carriers and is expanding into workers' compensation markets.1
In addition to his professional achievements, Vykruta has diverse interests and skills:
- He is a licensed commercial aircraft pilot and flight instructor.6
- He holds certifications in scuba diving and freediving.6
- He is multilingual, with proficiency in English, Czech, and working knowledge of Spanish, German, and French.6
Under Vykruta's leadership, EvolutionIQ has grown significantly, raising over $40 million in funding and achieving a valuation of over $200 million.12 Recently, in December 2024, EvolutionIQ was acquired by CCC Intelligent Solutions for $730 million, marking a significant milestone for the company.34
Highlights
Recursive Introspection: Teaching Language Model Agents How to Self-Improve is a new research paper on self-refining LLM's, incorporating error reasoning directly into Fine Tuning data for LLM's.
Here's a simplified breakdown of the RISE algo to achieve multi-turn correction:
-
Initial Data Collection:
- Collect a set of prompts and their corresponding desired outputs.
- Example: For a prompt "Translate 'hello' to French", the desired output is "bonjour".
-
Initial Model Training:
- Train the model on the initial dataset using standard supervised learning.
- The model learns to map prompts to outputs based on the provided examples.
-
Generate Responses:
- Use the trained model to generate responses for each prompt.
- Example: The model translates "Translate 'hello' to French" as "salut".
-
Identify Mistakes:
- Compare the generated responses to the desired outputs to identify mistakes.
- Example: "salut" is identified as incorrect since "bonjour" is the desired output.
-
Create Correction Prompts:
- For each incorrect response, create a new prompt asking the model to correct itself.
- Example: "The translation 'salut' is incorrect. Try again to translate 'hello' to French."
-
Generate Corrections:
- The model generates a new response based on the correction prompt.
- Example: The model might now generate "bonjour".
-
Feedback Loop:
- Continue this process iteratively, where each incorrect response leads to a new correction prompt.
- This creates a multi-turn dialogue, enabling the model to refine its responses over multiple attempts.
-
Fine-Tuning:
- Fine-tune the model using the collected correction prompts and their correct responses.
- The model learns to improve its initial responses based on the corrections provided.
-
Evaluation:
- Evaluate the model's performance on a separate validation set to ensure it can generalize the corrections.
-
Deployment:
- Deploy the improved model for real-world use, where it can continue to refine its responses based on user feedback.
This iterative fine-tuning process helps the model learn from its mistakes and improve its performance over multiple turns.
(thanks Andy C for sharing this) https://lnkd.in/gGtvrxPA
BREAKING -> Claude 3.5 Sonnet, Anthropic's latest model was just released, and it's significantly better than Google's Gemini 1.5 Pro Model.
By these benchmarks it's now outperforming OpenAI's GPT-4o in all but 2 categories.
spoiler: It's a poor math student.
It's significantly better than GPT4o at: -reasoning over text -code -graduate level reasoning
https://lnkd.in/ePZB8BNp

