Learn by Doing with Steven

Drone view Italy #26

1 month ago | [YT] | 0

Learn by Doing with Steven

Europe’s BEV market is moving beyond the early-adopter phase. In H1 2026, battery-electric cars reached 20.7% of new EU registrations, up from 15.6% a year earlier, with 1.22 million BEVs registered. [1] Yet most automotive comparison tools still present vehicles as specification tables or collapse them into opaque scores.

The more useful question is not “Which EV has the longest range?” but “How much range, efficiency and charging capability does each vehicle deliver for the money?”





Using a working sample of 99 BEV variants drawn from the European Alternative Fuels Observatory (EAFO), this note applies a Pareto-frontier framework to approximate EU price, range, battery capacity, efficiency and fast-charging speed. EAFO tracks EV models currently available to European consumers and publishes these core vehicle attributes. [2]

· Price and range remain strongly related, but a small group of vehicles materially outperform the normal price–range trade-off.

· Energy efficiency is more revealing than range alone: long range can be purchased with a larger battery, whereas efficiency reflects how effectively a vehicle converts stored energy into mobility.

· Tesla is unusually difficult to dominate simultaneously on price, range and efficiency in this sample.

· The newest Mercedes and Audi architectures materially change the road-trip frontier through very high charging productivity.

· BYD, MG and GWM exert strong pressure on acquisition price, but the current sample does not support a claim that Chinese OEMs dominate every technology dimension.

· The most important structural shift is that useful BEV capability is moving into lower price bands rather than only increasing at the premium end.



Learn By Doing With Steven 1 Learn By Doing With Steven Steven Auto Insight Steven Auto Insight 1 Steven Consultancy



#EU #EV #dataanalysis #learnbydoingwithsteven #stevenautoinsight #stevenconsultancy #paretofrontier

1 month ago | [YT] | 0

Learn by Doing with Steven

30B-class open models are getting close enough that a single benchmark score tells less and less of the story.

For a stricter comparison, I looked at three dense models in essentially the same parameter class:

• IBM Granite 4.2 30B
• Qwen3-32B — 32.8B
• DeepSeek-R1-Distill-Qwen-32B — 33B

This is a much cleaner comparison than mixing dense models with smaller models or MoE architectures that activate only a fraction of their parameters per token.

What stands out is how differently these models are being optimized.

**Granite 4.2 30B** is clearly targeting agentic and enterprise workloads. IBM reports 89.17 on AIME25, 75.77 on LiveCodeBench v6, 57.0 on SWE-Bench Verified and 61.39 on BFCL v4 — with particular emphasis on coding, tool use and multi-step agent workflows.

**Qwen3-32B** remains a strong general-purpose model, combining thinking and non-thinking modes with multilingual capabilities across 100+ languages. Its 32K native context can also be extended to 128K with YaRN.

**DeepSeek-R1-Distill-Qwen-32B** takes another route: transferring reasoning capabilities from DeepSeek-R1 into a much smaller Qwen2.5-based model. Its official evaluation reports 72.6 on AIME 2024 and 94.3 on MATH-500.

One important caveat:

These numbers come from official model cards, but they are **not one standardized leaderboard**. Benchmark versions, prompting and inference budgets differ across providers.

And that may be the more important point.

At the ~30B scale, the question is increasingly moving from:

**“Which model scores highest?”**

to:

**“Which model fits the workload best?”**

For enterprise AI, reasoning quality is only part of that decision. Tool reliability, context requirements, deployment control, licensing and governance increasingly matter just as much.

#OpenModels #EnterpriseAI #AgenticAI #LLM #AIGovernance #Granite #Qwen #DeepSeek

Skip the Carousel

1 2

1 month ago | [YT] | 0

Learn by Doing with Steven

🧠 The Brain in the Jar: Why AI Models Are Only Half the Equation

Have you ever wondered why two AI products running the exact same underlying model can feel completely different? One might be a basic back-and-forth chatbot, while the other functions as an autonomous developer tool capable of working for hours on complex software tasks.

The difference isn't the model itself. It comes down to the Agentic Harness.

Here is a breakdown of what actually drives AI performance and why the model is only part of the story.

1. The Model (The Brain) vs. The Harness (The Jar)

* The AI Model: Think of models like GPT-4 or Claude as a "brain in a jar." They are incredibly capable neural networks, but on their own, they are trapped. Natively, they cannot browse the web, open a local file, or execute code.
* The Agentic Harness: This is the system wrapped around the model. It provides the hands, eyes, and memory needed to interact with the world.
* The AI Agent: When you combine the AI Model and the Agentic Harness, you get a fully functional AI Agent.

2. Inside the Agentic Harness

A modern harness brings an AI to life using three core components:
* 🛠️ Tools (Interacting with the World): Tools allow the model to break out of its sandbox. This includes reading/writing files, executing code, searching the internet, or even driving a computer screen natively using computer use capabilities. Modern harnesses also leverage the Model Context Protocol (MCP) to seamlessly plug into external corporate databases and third-party apps.
* 💾 Memory (Managing the Context): Models have a fixed working memory (the context window) that wipes clean after a session. The harness steps in to persist crucial instruction files (like codebase conventions), summarize and prune old context to save space, and allow the model to search for specific code snippets rather than overwhelming it with an entire repository at once.
* 🔄 The Agentic Loop (Achieving Goals): This is the cycle of autonomy. The model formulates a plan, the harness takes an action, and the model observes the result. Modern harnesses continuously run verification throughout this loop—testing code and reviewing screenshots—ensuring the model stays on track during long-horizon tasks.

3. The Real Driver of AI Capabilities

As benchmark gaps between top-tier models shrink to within a few points of each other, the model itself is rarely the sole differentiator in product quality. Most of the recent massive capability gains in autonomous AI haven't just come from smarter brains; they’ve come from much better jars.

When evaluating if an AI is suited for a specific enterprise task, the question is no longer just "Which model are we using?" but rather, "Which harness are we wrapping it in?"

Reference: AI Model vs Agentic Harness: What Actually Drives AI by IBM Technology
Keep experimenting and building. Until next time on Learn by Doing with Steven.

1 month ago | [YT] | 0

Learn by Doing with Steven

人工智能智能体只阅读静态文档，学不会真实的专业工作。它需要进入能够复现任务世界、日常应用工具和结果验证机制的环境，在行动与反馈中训练。

这也是法律、金融、医疗等领域必须由专家设计评分量规的原因。这些任务往往没有唯一答案，环境既要奖励真实推理和工具使用，也要尽量阻止走捷径与奖励作弊。

阅读全文：learnbydoingwithsteven.github.io/
全部链接：linktr.ee/learnbydoingwithsteven

Ref: Sequoia Capital / Mercor / Brendan Foody

#LearnByDoingWithSteven #强化学习环境 #强化学习 #Mercor #智能体 #企业AI #人工智能训练 #机器学习

2 months ago | [YT] | 1

Learn by Doing with Steven

AI agents do not learn real professional work from static documents alone. They need environments that reproduce the world around a task, the applications used to complete it, and the verifiers that determine whether the result is acceptable.

This is why expert-built rubrics matter in law, finance, medicine, and other domains where there is rarely one perfect answer. The environment must reward genuine reasoning and tool use while making shortcuts and reward hacking harder.

Read the full analysis: learnbydoingwithsteven.github.io/
All links: linktr.ee/learnbydoingwithsteven

Ref: Sequoia Capital / Mercor / Brendan Foody

#LearnByDoingWithSteven #RLEnvironments #ReinforcementLearning #Mercor #AgenticAI #EnterpriseAI #AITraining #MachineLearning

2 months ago | [YT] | 0

Learn by Doing with Steven

软件越来越容易复刻，但一家企业的产品判断力仍然难以复制。这种判断力体现在产品推荐什么、拒绝什么、优先处理什么，以及怎样定义一个好结果。

后训练的价值，是把这些判断转化为稳定的模型行为。检索增强处理动态事实，监督微调稳定格式和边界，偏好训练吸收专家品味，强化学习培养多步骤能力，知识蒸馏则改善成本与延迟。

阅读全文：learnbydoingwithsteven.github.io/
全部链接：linktr.ee/learnbydoingwithsteven

Ref: Sequoia Capital / Fireworks.ai / Lin Qiao

#LearnByDoingWithSteven #后训练 #FireworksAI #企业AI #机器学习 #产品战略 #知识蒸馏 #强化学习

2 months ago | [YT] | 0

Learn by Doing with Steven

Software is becoming easier to reproduce, but a company's product judgment is still difficult to copy. That judgment lives in what the product recommends, rejects, prioritizes, and defines as a good result.

Post-training is the process of turning that judgment into model behavior. Retrieval handles changing facts; supervised fine-tuning stabilizes formats and boundaries; preference tuning captures expert taste; reinforcement learning develops multi-step capability; and distillation improves cost and latency.

Read the full analysis: learnbydoingwithsteven.github.io/
All links: linktr.ee/learnbydoingwithsteven

Ref: Sequoia Capital / Fireworks.ai / Lin Qiao

#LearnByDoingWithSteven #PostTraining #FireworksAI #EnterpriseAI #MachineLearning #ProductStrategy #ModelDistillation #ReinforcementLearning

2 months ago | [YT] | 1

Learn by Doing with Steven

“不是你的权重，就不是你的产品”点出了企业人工智能正在面对的问题。前沿模型接口仍然适合探索需求和验证产品市场契合，但完全依赖外部接口，会让企业承受推理成本、延迟、供应商变化和核心工作流控制不足等风险。

主权人工智能并不是一次性放弃所有外部模型。更务实的路径，是明确哪些能力可以租用、哪些能力必须拥有，再通过评测、执行框架、模型路由、后训练和反馈闭环建设自有智能层。

阅读全文：learnbydoingwithsteven.github.io/
全部链接：linktr.ee/learnbydoingwithsteven

Ref: Sequoia Capital / Sonya Huang

#LearnByDoingWithSteven #主权人工智能 #企业AI #开放权重 #后训练 #人工智能基础设施 #技术战略 #机器学习

2 months ago | [YT] | 0

Learn by Doing with Steven

"Not your weights, not your product" captures a growing enterprise AI concern. Frontier APIs remain excellent for exploration and product-market fit, but total dependence can expose the business to rising inference costs, latency constraints, provider changes, and weak control over proprietary workflows.

Sovereign AI is not an all-or-nothing migration. The practical strategy is to define what should be rented and what must be owned, then build the owned layer through evaluations, execution harnesses, routing, post-training, and feedback loops.

Read the full analysis: learnbydoingwithsteven.github.io/
All links: linktr.ee/learnbydoingwithsteven

Ref: Sequoia Capital / Sonya Huang

#LearnByDoingWithSteven #SovereignAI #EnterpriseAI #OpenWeights #PostTraining #AIInfrastructure #TechStrategy #MachineLearning

2 months ago | [YT] | 0