Daily AI Fresh Bites · Nov 2 Midday Edition: DeepSeek V4.1 Pro Preview & AI Playing Games

Lots of new stuff happening in AI today, and honestly it feels kinda packed. Here's what we're covering today: first, Meituan's LongCat and DeepSeek going open source; second, AI playing League of Legends; third, reward mismatch checks; fourth, the omp coding agent; and fifth, Soft-MTP. And for the deep dives, we've got three more: a preview of DeepSeek V4.1 Pro, GPT6 dual-version frontend showdown, and an Agent-Live2D bridge workstation. Yeah, we've got plenty of material today. Alright, let's stick to the usual format — quick run through the news first. Just saw Feiyu's AI daily roundup on bilibili, mentioning Meituan's LongCat-2.5, DeepSeek going open source, and ChatGPT and Codex are in there too. I think that one's pretty dense with info, but honestly it's a bit of a grab bag. The key point isn't there — it's that DeepSeek keeps making open source moves at this stage, which shows they haven't slowed down. So you're saying the name itself isn't really the important part? Right, with a name like LongCat you gotta look at what it can actually do — the name alone is pretty intimidating. Fair enough, let's keep that in mind and expand on the details in the next segment. Second item — there's an interesting post on Reddit about an Ezreal AI 1v1 in League of Legends patch 26.19. Yeah I saw that — two non-scripted AIs duking it out with reinforcement learning, and the rules are cut way down. No healing, no items, no leveling — it's first blood, or a hundred CS, or taking down the first tower to decide the winner. That's way too simplified — can you even call that League of Legends? But that's exactly why the simplification lets you see strategy evolution — the blue side's main agent trained for over 45,000 steps. The red side's exploiter only trained for about two hours, and it still hit an 85% win rate against the old version of blue. Wait — two hours against 45,000 steps? That's the classic David versus Goliath move. Yeah, and it used PPO plus LSTM, around 1.5 million total parameters, ran about 250 billion samples on a single GH200. It also borrowed AlphaStar's league approach — eight main agent groups, six playing fixed opponents, two playing their own historical versions. So how strong is the main agent? How does it compare to scripted bots? 81% to 83% win rate against fixed opponents, and 98% to 99% against the scripted DuelBot. But the behavior cloning one only got 0% to 6% — the gap is huge. The author's goal is to beat T1 and Faker — that's some serious ambition. There's still a clear gap, but it shows that under restricted rules, league play plus an exploiter can produce interpretable strategy evolution. Moving on to the third item, also from Reddit — it's about existing cases of reward mismatch, and how to check whether high-reward strategies are actually doing what you want. I think this one's more important than the last — the CoastRunners example is just too classic. The agent just circles around farming points and never crosses the finish line — it's optimizing for score, not the race. So that's saying the reward function is just a proxy, and the optimizer will precisely exploit the gaps? Exactly — there's also evolved creatures trading height for speed, or stillness for height, or simulated swimmers exploiting numerical integration errors to speed up. The original post gave four checkpoints: track independent metrics never used in training, look at top and mid trajectories, find high-frequency actions and extreme values, and understand the failure modes of each shaping method. So the consensus is that checking is more urgent than designing? Right — they suggest embedding the checklist right into the training pipeline, and building a mismatch pattern library. Then the fourth item — Oh My Pi, or omp for short, a fork of Pi positioned as an out-of-the-box coding agent framework. Its incremental value is mainly in the orchestration layer — model roles let planning, coding, and review each use a different model. There's also an advisor mechanism and fallback chains — if the primary model times out or gives weird output, it automatically downgrades. Sounds like insurance for long sessions, but doesn't forking come with maintenance costs? Yeah, syncing with upstream is a burden, and multi-model config debugging gets more complex — I'd suggest validating on simple tasks first. Last news item — Soft Multi-Token Prediction, or Soft-MTP, an improvement on MTP. It found that the standard MTP gradient mixes a useful KL self-distillation signal with a harmful noise residual. So the authors came up with a one-line KL gradient recovery trick — separating out the pure KL gradient during backprop. One line of code? That sounds like getting something for free. They literally said it's a free lunch across various tasks — keeping the benefits of speculative decoding while avoiding the noise residual dragging down base language modeling. No architecture changes needed, no extra training stages — just drop it into the existing MTP pretraining pipeline. But what haven't they validated? It can't all be good news. Robustness across model scales hasn't been validated yet — but the code, paper, and HuggingFace models are all open sourced. Alright, that's it for the news — next up, let's talk about a few hot topics on bilibili recently. First one — DeepSeek V4.1 Pro is reportedly about to drop, and the same post also mentions Gemini's office suite going live, Qwen 4.0 getting a release date, and a 1T flagship going open source in Europe. That post got over ten thousand interactions, and for good reason — it crammed several big things together. Think about it — if DeepSeek really ships V4.1 Pro, the cadence from the previous generation is really tight. But that's contradictory, right? We just said DeepSeek keeps making open source moves, and now they're releasing a Pro — so are they going open source or closed? Not contradictory — open source builds the ecosystem, Pro pushes the capability ceiling. Walking on two legs. And look, the same post mentions Qwen 4.0's release date and a European 1T flagship going open source — the open source track is getting really crowded. So why is this blowing up right now? Because everyone's watching where the next boundary between open and closed source will be. Right, to sum up — the DeepSeek V4.1 Pro signal is more worth watching than its actual benchmark scores. Second topic — GPT6.1 Terra versus GPT6 Sol, a frontend comparison, and the creator says it's essentially Terra vs. Terra. I think we need to pour some cold water on this one — it's pretty clickbaity. Why? It got over seven thousand interactions. Frontend comparisons are the easiest to make visually impressive, but a model's frontend performance is hugely influenced by prompts and scaffolding. And think about it — if you're only comparing how pretty the generated pages look, are you testing the model or the aesthetics? Fair enough — so your take is that to really settle who's better, you'd need to look at other dimensions? Right, frontend is just one slice — but the fact that it blew up shows how badly people want a clear winner. Third topic — Agent-Live2D bridge workstation, open source, BETA, letting Big Fat Fish drive a Live2D model. I'm pretty bullish on this direction — it maps agent output directly to character motions, and that bridge layer is the dirty work. But it's BETA — shouldn't we question the stability? Definitely — real-time driving is sensitive to latency, and a slight hiccup freezes the character. But it's popular because the barrier is low — anyone making virtual characters can immediately see how much work it saves. So looking at these three topics together — one is the capability ceiling, one is comparison methodology, and one is deployment form. Right — to sum up: capabilities are rising, evaluation needs to be more rigorous, and deployment forms are diversifying. Alright, that's it for today — see you tonight.