GenAI PMSebastian Raschka
AI researcher and educator sharing technical training notes on RLVR and GRPO. In the newsletter he is cited for a detailed reinforcement-learning-from-scratch thread.
Key Highlights
- Sebastian Raschka is repeatedly cited for detailed, from-scratch educational content on RLVR, GRPO, sampling, and LLM systems.
- His October 2026 thread covered GRPO mechanics including clipped policy ratios, KL loss, entropy, metrics, and format rewards.
- He shared practical guidance on inference scaling, reporting more than 2x higher answer accuracy via self-consistency and best-of-N techniques.
- His KV-cache explainer and memory calculator are useful references for understanding context-window and memory tradeoffs in product deployments.
- He consistently adds nuance to benchmarks, warning against overinterpreting calibration results, visual demos, or single-task performance gains.
At a glance
- Role
- AI researcher and educator
Overview
Sebastian Raschka is an AI researcher, educator, and technical communicator known for sharing detailed, implementation-oriented explanations of large language model systems and training methods. In the newsletter, he is cited repeatedly for from-scratch educational material on reinforcement learning with verifiable rewards (RLVR), GRPO, inference scaling, KV caches, sampling, watermarking, and model architecture details. He is also referenced through aliases including @rasbt and @Sebastian Raschka.
For AI Product Managers, Raschka matters because his work helps translate fast-moving research topics into practical mental models that are easier to evaluate for product planning. His posts and tutorials often focus on concrete mechanisms—such as clipped policy ratios, KL loss terms, verifier generation, calibration caveats, KV-cache memory tradeoffs, and sampling strategies—that directly affect model quality, reliability, latency, cost, and deployment decisions.
Key Developments
- 2026-08-12: Recapped Meta’s release of Muse Glimmer, describing it as an open-weight dense 30B multimodal reasoning model with a 131k context window, and highlighted its memory-efficient KV-cache design for agentic workflows.
- 2026-08-13: Shared links to Meta AI’s Muse Glimmer announcement and the Hugging Face page for Muse-Glimmer-30B, emphasizing that the release was real.
- 2026-08-16: Explained watermarking as an additional selection mask applied at certain positions during top-p or top-k sampling, with a seed dependent on the previous token; also noted that watermarking can be implemented without changing the raw distribution by making the sampling seed depend on the previous token using a private key.
- 2026-08-31: Commented on dependency pinning and CUDA compatibility, noting that adding a `uv.lock` can help somewhat, but excessive pinning can create issues because a version that works for most users may fail for others due to driver compatibility.
- 2026-09-07: Shared a from-scratch article on KV caches in LLMs and an interactive memory calculator configured for Qwen3-0.6B, illustrating token-memory tradeoffs.
- 2026-09-14: Shared “Reasoning from scratch round 3,” focused on verifier generation for evaluating a base model and preparing later RLVR workflows, with discussion of evaluation, reproducibility, prompt sensitivity, and base versus reasoning models.
- 2026-09-16: Demonstrated a Paint UI benchmark comparing GPT-5.6 Astra and Qwen3.8 Max, while cautioning that a better-looking final image alone does not prove stronger generalization, computer-use ability, or visual understanding.
- 2026-09-20: Shared “Inference scaling part 1,” using modified text generation with temperature scaling, top-p filtering, and multinomial sampling to produce diverse outputs for self-consistency and best-of-N; he reported more than 2x higher answer accuracy and discussed MATH-500 compute tradeoffs and self-refinement.
- 2026-09-21: Commented that a model being right 70% of the time when it says “70% likely” shows calibration only on the evaluated dataset, not necessarily on out-of-distribution data.
- 2026-10-11: Shared round 2 of “Reasoning From Scratch: Reinforcement Learning with Verifiable Rewards (RLVR),” covering GRPO details such as clipped policy ratios, a KL loss term, format rewards, entropy, metrics, and other training tips.
Relevance to AI PMs
1. Useful for evaluating model training claims: Raschka’s RLVR and GRPO explanations help AI PMs ask better questions about how a reasoning model was improved, what reward signals were used, and whether reported gains come from training, inference scaling, or evaluation setup.
2. Helpful for cost-performance planning: His KV-cache, architecture, and inference-scaling breakdowns give PMs concrete ways to reason about latency, memory footprint, throughput, context-window costs, and when techniques like best-of-N or self-consistency may be worth the spend.
3. Improves benchmark and reliability judgment: His comments on calibration, verifier generation, reproducibility, and benchmark interpretation are directly useful when deciding whether a demo, leaderboard gain, or visual benchmark result is product-relevant versus merely attention-grabbing.
Related
- GRPO / RLVR / reinforcement-learning-with-verifiable-rewards: Central to Raschka’s recent “Reasoning from scratch” materials, especially around verifier-based training and reward design.
- KV caches / kv-caches / qwen3-06b: Connected through his from-scratch explainer and memory calculator work on token-memory tradeoffs in LLM inference.
- Inference scaling / MATH-500 / reasoning-models: Linked through his post on self-consistency, best-of-N, and accuracy-compute tradeoffs.
- Top-p sampling / top-k sampling / watermarking: Related to his explanations of sampling mechanics and watermarking implementation choices.
- Meta AI / Muse Glimmer / open-weight-models: Connected via his commentary on Meta’s multimodal reasoning release and its memory-efficient architecture.
- Transformers / attention / grouped-query-attention / multi-head-latent-attention: Relevant because his educational content often breaks down architecture-level tradeoffs important for model selection.
- Hugging Face / GitHub / llms-from-scratch / reasoning-from-scratch: These connect to the distribution and discoverability of his educational resources and code-oriented explanations.
- Simon Willison, Philipp Schmid, Jeff Dean, Demis Hassabis, Nathan Lambert, Nato Lambert: Adjacent voices in the broader technical discussion around LLM systems, research communication, and model evaluation.
Frequently asked questions
What is Sebastian Raschka?
AI researcher and educator sharing technical training notes on RLVR and GRPO. In the newsletter he is cited for a detailed reinforcement-learning-from-scratch thread.
What should AI product managers know about Sebastian Raschka?
Sebastian Raschka is repeatedly cited for detailed, from-scratch educational content on RLVR, GRPO, sampling, and LLM systems. His October 2026 thread covered GRPO mechanics including clipped policy ratios, KL loss, entropy, metrics, and format rewards. He shared practical guidance on inference scaling, reporting more than 2x higher answer accuracy via self-consistency and best-of-N techniques.
Explore this topic
Newsletter Mentions (54)
“#5 𝕏 Sebastian Raschka shared round 2 of “Reasoning From Scratch: Reinforcement Learning with Verifiable Rewards (RLVR),” covering clipped policy ratios, a KL loss term, format rewards, and other GRPO tips and tricks.”
#5 𝕏 Sebastian Raschka shared round 2 of “Reasoning From Scratch: Reinforcement Learning with Verifiable Rewards (RLVR),” covering clipped policy ratios, a KL loss term, format rewards, and other GRPO tips and tricks. The original post includes detailed chapter timestamps on GRPO training, metrics, entropy, policy ratios, KL divergence, and format rewards. Also covered by: @Sebastian Raschka
“Sebastian Raschka commented that a system saying “70% likely” and being right “70% of the time” demonstrates calibration only on a specific dataset, not guaranteed calibration on out-of-distribution data.”
Sebastian Raschka commented that a system saying “70% likely” and being right “70% of the time” demonstrates calibration only on a specific dataset, not guaranteed calibration on out-of-distribution data.
“Sebastian Raschka shared “Inference scaling part 1,” which uses a modified text-generation function with temperature scaling, top-p filtering, and multinomial sampling to generate diverse outputs for self-consistency and best-of-N.”
#4 𝕏 Sebastian Raschka shared “Inference scaling part 1,” which uses a modified text-generation function with temperature scaling, top-p filtering, and multinomial sampling to generate diverse outputs for self-consistency and best-of-N. He reports more than 2x higher answer accuracy and covers MATH-500 results, compute tradeoffs, and self-refinement.
“Sebastian Raschka demonstrated a Paint UI benchmark in which GPT-5.6 Astra layered geometric shapes while Qwen3.8 Max worked pixel by pixel, producing a closer match. He cautioned that final-result similarity alone does not establish better generalization, computer-use capabilities, or visual understanding.”
#11 𝕏 Sebastian Raschka demonstrated a Paint UI benchmark in which GPT-5.6 Astra layered geometric shapes while Qwen3.8 Max worked pixel by pixel, producing a closer match. He cautioned that final-result similarity alone does not establish better generalization, computer-use capabilities, or visual understanding. #12 𝕏 Chip Huyen commented on an approach where models choose from predefined values instead of generating freeform text, potentially benefiting data labeling and fixed-action tasks.
“𝕏 Sebastian Raschka shared “Reasoning from scratch round 3,” covering verifier generation for evaluating a base model against future improvements and for later reinforcement learning with verifiable rewards (RLVR).”
#3 𝕏 Sebastian Raschka shared “Reasoning from scratch round 3,” covering verifier generation for evaluating a base model against future improvements and for later reinforcement learning with verifiable rewards (RLVR). Timestamped chapters address LLM evaluation, verifiers, MATH-500, prompt sensitivity, reproducibility, and base versus reasoning models. Also covered by: @Sebastian Raschka
“Sebastian Raschka shared his from-scratch article about KV caches in LLMs.”
#5 𝕏 Sebastian Raschka shared his from-scratch article about KV caches in LLMs. #9 𝕏 Sebastian Raschka shared an interactive memory calculator hosted on sebastianraschka.com, configured for Qwen3-0.6B with 8192 tokens, batch=1, weight=16, and kv=16.
“He noted that some people cannot run a specific version because of CUDA compatibility with specific drivers, and that the default works for most people, but not everyone.”
#6 𝕏 Sebastian Raschka said he added a `uv.lock` to help a little but cautioned that excessive pinning can cause issues. He noted that some people cannot run a specific version because of CUDA compatibility with specific drivers, and that the default works for most people, but not everyone.
“Sebastian Raschka described watermarking as an additional selection mask at certain positions during top-p/k sampling. In his example, it randomly samples among 5 likely top-p choices using a seed dependent on the previous token.”
#9 𝕏 Sebastian Raschka described watermarking as an additional selection mask at certain positions during top-p/k sampling. In his example, it randomly samples among 5 likely top-p choices using a seed dependent on the previous token. #10 𝕏 Sebastian Raschka said changing the distribution is unnecessary: making the random seed depend on the previous token using a private key can enable specific, reproducible top-p or top-k sampling without altering the raw distribution.
“Sebastian Raschka shared links to Meta AI’s introduction of Muse Glimmer and the `meta-models` Hugging Face page for Muse-Glimmer-30B, emphasizing that it is real—not an April 1st joke.”
#2 𝕏 Sebastian Raschka shared links to Meta AI’s introduction of Muse Glimmer and the `meta-models` Hugging Face page for Muse-Glimmer-30B, emphasizing that it is real—not an April 1st joke. Also covered by: @Fireship
“"#4 𝕏 Sebastian Raschka recaps Meta’s release the previous day of Meta Muse Glimmer, an open-weight, dense 30B multimodal reasoning model with a 131k context window."”
#4 𝕏 Sebastian Raschka recaps Meta’s release the previous day of Meta Muse Glimmer, an open-weight, dense 30B multimodal reasoning model with a 131k context window. Its 32 query heads and 2 KV heads deliver a 52 KiB BF16 KV cache per token, making it memory-efficient and well suited to agentic workflows, though benchmarks conflict on whether it outperforms Qwen3.6. Also covered by: @Rowan Cheung , @Rowan Cheung
Related
An AI company mentioned for releasing a report on model behaviors and for internal tool use in later context. Relevant for PMs tracking safety reporting and frontier model operations.
Anthropic’s coding agent/environment used here to control physical tools in a long-running home experiment.
OpenAI is an AI company building foundation models and consumer products like ChatGPT. In this newsletter it is discussed in connection with GPT-6, Dots, ChatGPT for Teens, and College Planner.
An OpenAI coding model/tool used here as part of coding harnesses turned into reinforcement-learning environments.
AI practitioner credited here with highlighting the public availability of SynthID Detector.
A prominent AI blogger and commentator referenced in connection with an article on token reselling and fraud. He is cited as the source of the newsletter item discussing the marketplace and API-key abuse.
An AI model hub and community platform referenced through comments about its usage and download metrics. The newsletter suggests DAUs are more reliable than raw download counts for understanding growth.
Google's advanced AI research organization. In this newsletter it is mentioned via SynthID Detector and its Gemini live models.
OpenAI’s chat product, referenced here as the user-owned subscription plan connected to Devin. For PMs, this highlights quota-linked integration patterns.
Alibaba's AI model and agent ecosystem brand. In this newsletter it announced new audio models, agent capabilities, and benchmark releases.
Technology company referenced as the place where John Provine previously built evaluation systems. Relevant here as a source of product experimentation rigor.
A major AI company involved here in developing the open Personal Agent Protocol with Sierra. The mention is relevant to interoperability and agent permissions.
CEO and co-founder of Google DeepMind, mentioned here announcing open-source SynthID Bio tools.
A Google AI leader cited commenting on Waymo safety-data improvements. Relevant to PMs as a prominent voice on AI progress, benchmarking, and real-world reliability.
An AI discovery platform that demonstrated a motion-graphics video generated from a prompt about prompt injections. For PMs, it illustrates how AI tools are used to explain security concepts through generated media.
A model used as an automated judge in Claire Vo’s benchmark. It contributes 30% of the scoring alongside her manual evaluation.
A model family discussed in the context of technical architecture and inference efficiency. The report highlights attention design, KV cache reduction, and faster decoding methods.
A Claude model variant being updated with stronger biology safeguards to reduce false positives while still routing dual-use biology requests to higher-safety fallback behavior. Relevant for PMs considering safety tradeoffs and product-surface-specific policy tuning.
The code-hosting and collaboration platform used for pull requests and repository workflows. Here it is part of Warp’s automated engineering loop.
A model routing and access platform where MAI-Transcribe-2 reportedly reached 1 million requests. It is relevant as a distribution and usage layer for AI models.
A 2.8T-parameter open-weight model described as frontier-level by the speaker in the newsletter. It is notable for strong quality and deployment on Nebius Token Factory.
A Grok-related product or builder experience referenced as part of the SpaceXAI stack. It appears in the context of improvements from Cursor’s integration.
Meta’s AI organization behind model and product releases. PMs should note it as the source of Muse Glimmer and the associated Hugging Face release.
Agents used to write, review, and iterate on code as part of software development workflows. The newsletter frames them as shifting developers toward specification, architecture, and evaluation work.
A large language model used as the reasoning core inside agents and tool-calling systems. PMs often evaluate LLMs based on orchestration, context loading, and task execution behavior.
A local LLM runtime and API server for running open models on personal hardware. The newsletter includes commands and memory guidance for using it with Gemma locally.
Large language models are referenced as capable of writing essays but limited in physical task learning and control. The newsletter uses them as a baseline for comparing future architectures.
A Qwen model release with day-0 support for multimodal integration. The newsletter highlights its immediate compatibility with MLX-VLM for visual-language workflows.
An AI company or model provider mentioned for releasing a cost-efficient open-weights multimodal model. For PMs, it signals open-weight multimodal competition and cost pressure.
A model referenced in the newsletter’s overview of recent LLM architectures. It appears here as an example of architecture-level innovation and efficiency work in foundation models.
A gallery or reference resource used to compare LLM architectures and models. It is referenced as the place where Qwen3.6 and Kimi-K2-6 are compared.
Research scientist and podcaster focused on AI, robotics, and technical conversations. Here he announces a long-form technical AI podcast spanning training architectures, robotics, compute, business, and geopolitics.
AGI refers to broadly capable artificial general intelligence. Here it is discussed as becoming usable in 2026 and requiring contextual systems around it to be effective.
Apple's on-device AI layer powering features like Live Translation on supported hardware. Relevant to PMs as part of Apple’s AI product stack and device-gated rollout.
An agent design pattern where work is split into sub-tasks and assigned dynamically. In the newsletter, it is one of the core ingredients for building autonomous coding agents.
A memory architecture pattern for AI agents that separates different memory layers to improve context retention and task performance. It is presented as part of the design of autonomous coding assistants.
The practice of connecting agents to external developer tools such as linters and debuggers. It is highlighted here as a building block for effective coding agents.
Stay updated on Sebastian Raschka
Get curated AI PM insights delivered daily — covering this and 1,000+ other sources.
Subscribe Free