Alex Volkov moved from full-stack software engineering and an AI startup toward public teaching because sharing what he learned became more rewarding than building the product itself.
2
Multimodal systems make both inference and evaluation harder because models must handle images, video, motion, audio, and possibly other signals while the supporting datasets and tools catch up.
3
AI agents are improving as context windows grow, but Alex still has not found one that can reliably manage the research and show-note workflow for ThursdAI.
Summary
Alex Volkov describes how a long career in generalist software engineering led him into AI evangelism. DALL-E, Stable Diffusion, and Whisper pulled him into model experimentation, and a weekend project became Targum Video, a service that translated videos on X. He then found that writing and speaking publicly about AI gave him more energy than adding features to the startup. That work became ThursdAI, a weekly podcast, newsletter, and live discussion. The conversation covers multimodal models, video embeddings, evaluation, retrieval systems, and agents. Alex argues that production AI requires evaluation of the whole system, including retrieval, embeddings, speed, cost, hallucinations, and user feedback. He gives practical lessons from a Weights & Biases hack week: define the use case, collect feedback, automate data ingestion, and build evaluations that allow model and embedding changes to be compared. On agents, he is interested in their progress but remains cautious because task completion and evaluation are still unreliable.
Alex entered AI through creative tools and open model weights
DALL-E changed Alex's view of AI because it gave him a capability he did not have, drawing. Stable Diffusion then exposed him to a different model experience: a weights file, some Python code, and the freedom to run it himself. Whisper brought him further into practical systems. He used it to translate a video from Ukraine and add subtitles, then built an X bot that automatically translated videos. That project became his startup for about seven months and forced him to learn how to put models into production, scale them, and monitor them.
Public learning became more satisfying than building the startup
Alex found that his strongest learning process was explaining new ideas publicly. He used X as a place to record what worked and what failed, and those practical notes attracted people who were learning the same things. After ChatGPT appeared, he built side projects to understand its conversational memory, context handling, and role-playing abilities. Over time, he noticed that discussing AI innovation was more exciting than adding features to Targum Video. A recurring Thursday conversation about GPT-4 became ThursdAI, then expanded into a podcast and newsletter with guests and regular segments.
Alex expects multimodality to be a major area of AI work. He describes models that combine language with visual understanding, and points out that other systems can use EEG and fMRI signals as inputs. Meta has explored models involving text, images, video, and IMUs, which provide information about motion. Alex connects this to robotics, where a system needs to understand how scenes and objects change over time. He also distinguishes ordinary audio transcription from richer audio understanding, since speech converted into text loses details such as intonation.
Video systems create large inference and retrieval problems
Video adds scene boundaries, cuts, motion, audio, storage, and hardware demands to an already difficult problem. Treating every video frame as an image creates a large amount of input, and Transformer attention grows quadratically with the number of tokens. Alex mentions 12 Labs as an example of work on video embeddings and video foundation models. For retrieval systems, he says that taking a screenshot from every frame is not a satisfactory way to preserve the context of a long movie. Systems need better ways to represent scenes and retrieve the information that matters.
Evaluation has to cover the model, system, and user outcome
Alex says model evaluation is already difficult because teams compare systems using task-specific benchmarks such as MMLU and GSM8K. Multimodal systems make the problem harder because evaluation may need to cover text, images, video, speech, and emotion. He also wants more general evaluations, since a model can beat another model on one narrow task while being less useful overall. In production, evaluation includes training, retrieval, generated output, truthfulness, hallucinations, cost, speed, and how well the system follows instructions. Demetrios Brinkmann adds that retrieval systems introduce further questions about evaluating the retriever and each component of the larger system.
A production RAG bot needs ingestion, feedback, and repeatable tests
Alex's Weights & Biases build-week examples show why a RAG prototype is different from a maintained product. It is easy to download documents, chunk them with LangChain or LlamaIndex, add embeddings, and get answers that look useful. The harder work is defining the use case, collecting thumbs-up and thumbs-down feedback, teaching users to provide that feedback, and storing enough data to rerun tests. Documentation changes, so a mature bot needs an ingestion pipeline that updates automatically. Once examples and user feedback exist, an evaluation pipeline can compare models, embeddings, speed, and hallucination rates.
The user experience determines which trade-offs are acceptable
Alex says teams should decide what the user is waiting for before choosing an architecture. A chatbot may need fast responses because users are accustomed to GPT-4-like interaction. A documentation translation job can run more slowly because nobody is waiting for the answer in real time. That difference changes the balance between model quality, cost, latency, and the number of prompts used. He also describes perceived latency. His Visual Weather GPT generates a weather image, but it first writes a poem so users have something to read while browsing and image generation take place. The backend did not become faster, but the waiting experience improved.
Agents are advancing, but reliability and evaluation remain weak
Alex traces the excitement around agents from AutoGPT and says many users quickly encountered hallucinations and failures to complete tasks. He sees progress from longer context windows and frameworks such as CrewAI, where separate agents can handle separate tasks. That structure also makes evaluation more manageable because each agent has a defined job. Still, Alex has not found an agent that can reliably gather his saved X posts and turn them into ThursdAI show notes. He treats that personal workflow as a useful test and remains unconvinced that agents are ready to act as a general solution.
