On the independent test set here, the accuracy goes from 15% (base model) to 48% (reasoning model). So yes, there is a big gain to be had. But of course, this is still far from larger, state-of-the-art models that can't run on laptops. I.e. those have >90% accuracies on the same dataset. So, overall, the model used in this video series is for education purposes.
That 15 to 48 percent jump is exactly what makes the topic click for students, and keeping it honest about the SOTA gap is the right call. When a smaller model runs on a normal laptop, do you find the training loop itself starts shaping the reasoning behaviour, or is the gain mostly in how you score and filter the outputs at inference time
I am trying to run the code on my PC CPU and I am stuck, I have tried with Lighting AI buy It Is not possibile to create a new Studio leveraging GPU are there any other alternatives, I am at chapter 5 of the book and It Is amazing . Thanks for suggestions
Hm, is this a platform issue? Have you tried to reach out to their customer support perhaps? Otherwise, Lambda Labs (https://lambda.ai would be a good alternative. I’ve used it in the past, and while the UI is not quite as nice, it works quite well, too.
Love that the verifier gets its own lecture. Most people treat evaluation as the boring bit before RL, when it's really the part that decides what you end up training.
Thank you for recording these, 11 hours of from-scratch code is a generous gift. Starting from a pre-trained Qwen3 0.6B and building the verifier yourself is the part I appreciate most: having the same verifier do double duty for MATH-500 evaluation and as the RL reward makes it very concrete why answer extraction and normalization matter so much. Looking forward to working through the distillation part on my homelab GPU. Did you find 0.6B large enough to show clear gains from RLVR, or do the effects get noisy at that scale?
I recently read your article, *Building Reasoning Models from the Ground Up*, and I am interested in learning how to conduct AI research independently and eventually contribute original research.
I have a BE in Electronics and Communication Engineering, and I am interested in LLMs, deep learning, computer graphics, and GPU computing. However, I have a serious concern about pursuing a PhD.
Many engineering graduates invest years in education, yet struggle to find suitable employment. I am also concerned that some PhD students may not receive adequate research mentorship or practical guidance on identifying research gaps, conducting experiments, writing papers, and publishing through venues such as arXiv, IEEE conferences, or journals.
**In your opinion, how can an aspiring researcher determine whether pursuing a PhD is a worthwhile and reasonably safe career decision, rather than spending several years without adequate mentorship or clear career prospects?**
Would you recommend first learning research methodology, reproducing published results, and developing a small research portfolio independently before applying for a PhD? What signs should prospective students look for to identify a good research supervisor and a research environment that genuinely prepares them for both academic and industry careers?
I would greatly appreciate your advice on how someone from my background, with limited financial resources, can pursue research responsibly without taking an unreasonable financial or career risk.
Thank you for sharing your knowledge and making complex AI research more accessible to independent learners.
Inference-time scaling looks cheap in a notebook. In production, 8 samples per request means roughly 8x the output tokens on that path, so we put a per-request budget on it before it ships.
The engine-versus-harness framing is the cleanest way I've seen this explained, and it maps onto money decisions more than people expect.
I keep watching clients pay for the harness — extra SIP top-ups, complex products, one more percentage point of certainty — while the base model is still broken: no emergency fund, no term cover, equity allocation guessed at.
Inference scaling only pays off once the base model is sound. Households work the same way.
The side note is the part I would keep next to the eight videos. A reasoning model is the engine after reinforcement learning with verifiable rewards and optional inference scaling, and the harness is what manages context and hands the model tools for a booking or a longer coding task. Variable reasoning effort can make the engine think longer and still miss a tool result the harness never put back into the next prompt. I would log the observations the harness actually returned, next to the effort setting, so a failed task is not scored as if the model never saw the tool output.
Framing the model as the engine and the agent harness as the machinery is a useful line. Most coverage blurs the two. I work on Jyst, which summarizes AI coverage and rates claims from hype to reality. This came out Holds up: the code is linked and it runs on an open Qwen3 0.6B base, so readers can check every step themselves. https://jyst.ai/stories/building-reasoning-models-from-the-ground-up-a-video-course If the summary misses anything, I'll fix it.
Reasoning models explained from scratch is what students need. Do you think a small model can show real reasoning gains on a normal laptop?
On the independent test set here, the accuracy goes from 15% (base model) to 48% (reasoning model). So yes, there is a big gain to be had. But of course, this is still far from larger, state-of-the-art models that can't run on laptops. I.e. those have >90% accuracies on the same dataset. So, overall, the model used in this video series is for education purposes.
That 15 to 48 percent jump is exactly what makes the topic click for students, and keeping it honest about the SOTA gap is the right call. When a smaller model runs on a normal laptop, do you find the training loop itself starts shaping the reasoning behaviour, or is the gain mostly in how you score and filter the outputs at inference time
Thank you for the amazing content, as always. Now i have something to look forward to during the holiday season.
Love the picture at the end. Happy regeneration.
I am trying to run the code on my PC CPU and I am stuck, I have tried with Lighting AI buy It Is not possibile to create a new Studio leveraging GPU are there any other alternatives, I am at chapter 5 of the book and It Is amazing . Thanks for suggestions
Hm, is this a platform issue? Have you tried to reach out to their customer support perhaps? Otherwise, Lambda Labs (https://lambda.ai would be a good alternative. I’ve used it in the past, and while the UI is not quite as nice, it works quite well, too.
Love that the verifier gets its own lecture. Most people treat evaluation as the boring bit before RL, when it's really the part that decides what you end up training.
Ah yes, it’s a crucial part here (and gets reused in RLVR)
Thank you for recording these, 11 hours of from-scratch code is a generous gift. Starting from a pre-trained Qwen3 0.6B and building the verifier yourself is the part I appreciate most: having the same verifier do double duty for MATH-500 evaluation and as the RL reward makes it very concrete why answer extraction and normalization matter so much. Looking forward to working through the distillation part on my homelab GPU. Did you find 0.6B large enough to show clear gains from RLVR, or do the effects get noisy at that scale?
Yeah the improvement was from about 15% to 48% accuracy on MATH-500
Thank you for these wonderful resources!
Amazing. Truly amazing. Thank you.
Hello Dr. Raschka,
I recently read your article, *Building Reasoning Models from the Ground Up*, and I am interested in learning how to conduct AI research independently and eventually contribute original research.
I have a BE in Electronics and Communication Engineering, and I am interested in LLMs, deep learning, computer graphics, and GPU computing. However, I have a serious concern about pursuing a PhD.
Many engineering graduates invest years in education, yet struggle to find suitable employment. I am also concerned that some PhD students may not receive adequate research mentorship or practical guidance on identifying research gaps, conducting experiments, writing papers, and publishing through venues such as arXiv, IEEE conferences, or journals.
**In your opinion, how can an aspiring researcher determine whether pursuing a PhD is a worthwhile and reasonably safe career decision, rather than spending several years without adequate mentorship or clear career prospects?**
Would you recommend first learning research methodology, reproducing published results, and developing a small research portfolio independently before applying for a PhD? What signs should prospective students look for to identify a good research supervisor and a research environment that genuinely prepares them for both academic and industry careers?
I would greatly appreciate your advice on how someone from my background, with limited financial resources, can pursue research responsibly without taking an unreasonable financial or career risk.
Thank you for sharing your knowledge and making complex AI research more accessible to independent learners.
Best regards,
Gobal Krishnan V.
Inference-time scaling looks cheap in a notebook. In production, 8 samples per request means roughly 8x the output tokens on that path, so we put a per-request budget on it before it ships.
The engine-versus-harness framing is the cleanest way I've seen this explained, and it maps onto money decisions more than people expect.
I keep watching clients pay for the harness — extra SIP top-ups, complex products, one more percentage point of certainty — while the base model is still broken: no emergency fund, no term cover, equity allocation guessed at.
Inference scaling only pays off once the base model is sound. Households work the same way.
The side note is the part I would keep next to the eight videos. A reasoning model is the engine after reinforcement learning with verifiable rewards and optional inference scaling, and the harness is what manages context and hands the model tools for a booking or a longer coding task. Variable reasoning effort can make the engine think longer and still miss a tool result the harness never put back into the next prompt. I would log the observations the harness actually returned, next to the effort setting, so a failed task is not scored as if the model never saw the tool output.
Framing the model as the engine and the agent harness as the machinery is a useful line. Most coverage blurs the two. I work on Jyst, which summarizes AI coverage and rates claims from hype to reality. This came out Holds up: the code is linked and it runs on an open Qwen3 0.6B base, so readers can check every step themselves. https://jyst.ai/stories/building-reasoning-models-from-the-ground-up-a-video-course If the summary misses anything, I'll fix it.