Daily AI Fresh News · October 29 Evening Edition: Training LLaMa from Scratch and Agent Testing Buzz

Lots of new stuff in the AI world today — someone on Reddit built a large model from scratch, and Agent tutorials are blowing up on Bilibili. Here's what we're covering today: first, hand-training LLaMa; second, new model benchmarks; third, Reap for expense reports; fourth, RAG production pitfalls; fifth, AI video dubbing. For deep dives, we'll talk about the WorkBuddy office Agent, how to break into Agent testing roles, and the CLI-Anything command-line tool. Yeah, these are all pretty representative. Let's start with the news roundup and go through them one by one. First up, there's a Reddit post called 'I trained LLaMa from scratch' — fully hand-written reproduction. I saw that one. The point isn't the result — it's the process. He didn't use any off-the-shelf framework — RoPE, GQA, RMSNorm, SwiGLU all hand-written, even the KV cache batch inference was built from scratch. So basically an entire architecture course put into practice from start to finish. Exactly, and the training strategy was very pragmatic — started with tiny-shakespeare for smoke testing. Then trained a 41M parameter model on TinyStories-v2 to validate generation, and finally a 125M model on FineWeb with 5.27B tokens. Two A5000s running for 11 hours — that's pretty affordable for individual developers. He said the next step is RLHF, but that needs multi-model pipelines for memory — probably needs some careful thought. I think the value of this kind of project is building engineering intuition — it's not aimed at being a product. Exactly. Understanding architecture trade-offs is way more valuable than just calling APIs. Second one — there's a YouTube video with a clickbaity title: 'Gemini 4 Carbon,' complete with Fable 5.5 benchmarks. Definitely clickbait vibes, but these benchmark videos do get a lot of traffic these days. It pits Gemini 4 Ultra against Claude Ultrathink — what do you think of these head-to-head comparisons? Fun to watch, but benchmarks are far from real tasks — don't use them as a basis for choosing a model. Third one — also on YouTube — Reap MCP, using an AI agent to automate business expense reports. That's a solid direction — expense reporting is a typical workflow task, perfect for an agent to take over. The key is it uses MCP, which basically plugs financial systems into the agent ecosystem. Right, but permissions need to be locked down tight — otherwise the agent submitting random expense reports would be a mess. Fourth — there's a workshop preview on the LangChain subreddit about RAG production pitfalls. I totally agree with this — demo tutorials only show you the happy path. In real production, tool call timeouts, retry loops, excessive permissions — it's all landmines. It says you need structured tracing to debug wrong answers, plus fault injection testing. Right, the focus is really on the engineering infrastructure around the model, not the model itself. And it runs on a laptop — no GPU, no API key needed — super low barrier to entry. Last one — GlobalCast AI v3.2, for AI video dubbing, unlimited and zero credits. It's BYOK — bring your own Gemini key plus MCP server — shifting the cost to users. That's a clever model — platform doesn't burn money, users bear their own token costs. Dubbing is a huge demand — content going global relies on it. Alright, that's the news. Now let's dive into some hot topics that have been blowing up on Bilibili recently. First one — the WorkBuddy tutorial, titled 'The most powerful office automation Agent, 99% of people underestimate it.' That one has high engagement — over 1,800 — which shows people are really looking for practical use cases. It focuses on three real-world case studies, step by step. Where do you think office Agents are right now? My take: single tasks are doable, but cross-system workflows are still a ways off. The trouble with office scenarios is scattered data and messy permissions — agents easily get stuck in the middle. But the fact that these tutorials are popular doesn't that show there's strong demand? Demand is definitely strong, but popularity doesn't mean the product is mature — a lot of it is anxiety-driven. Makes sense — like how everyone rushed to learn prompt engineering a while back. Exactly. Tools change, but the underlying approach of breaking down tasks and defining boundaries stays the same. So the conclusion is: use office Agents as assistants for now, don't expect them to fully take over automatically. Second topic is even more interesting — '2026's hottest AI testing job: zero-experience entry into agent testing.' I'm all for this — agent evaluation is a rare skill right now. Why is it suddenly so hot? Weren't testing roles pretty marginal before? Because agent outputs are non-deterministic — traditional assertions just can't test them. You have to design scenarios, define scoring criteria, and guard against it falsely claiming completion. False claims are really a killer — didn't some tests show models often lie about completing tasks? Exactly, so evaluation needs to look at actual database changes, not trust what it says. So is it realistic to transition in with zero experience? I feel like just knowing tools isn't enough. Not enough — you need product thinking plus some engineering foundation, otherwise you can't design good test cases. Fair enough — so the bar is actually not low, it's just packaged as zero-experience. To sum up: agent testing is hot because evaluation has become a bottleneck for deployment — this job is very valuable in the short term. Last one — CLI-Anything, letting Claude Code control Blender and other arbitrary software. Everything can be a command line — I really like this approach. How does it do it? The software doesn't have a ready-made CLI. The core idea is wrapping the software's operation interface as a command line so the agent can call it. It's like putting a programmable shell around each application, bypassing the GUI. So that's basically extending the agent's reach into all desktop software? Right, but stability depends on the quality of the wrapper — something complex like Blender can easily break. But I think the direction is right — if an agent is going to get work done, it needs to operate real tools. Exactly. So the conclusion is: CLI-ification is the most realistic path for agents to integrate with traditional software. From today's chat — from hand-building models to agent deployment — it feels like the whole field is moving toward engineering details. Yeah, it's no longer about whose model is bigger — it's about who can get the job done reliably. Alright, that's it for today. See you tomorrow.