Tags / Model research
LLM
- Is selling software a fake problem now? The thread says the fake problem is that question A V2EX post cites Meng Yan: if everyone can vibe-code an app, selling apps gets harder. 36 replies mostly disagree—people can cook; restaurants still exist. Same ruler as “software isn’t valuable”: the shell below the kill line depreciates, not the whole business of selling.
- Claude’s Double Standard: Ingest Copyright to Train, Guard Studio IP on Output After Anthropic barred Claude from recreating existing games, Reddit’s top takes weren’t about compliance—they were about whose IP earns a system-prompt.
- Do Chinese people still need English in the AI era? I want outsiders to weigh in A Chinese parent asked on V2EX whether kids should still grind English now that AI can translate anything. Almost nobody said “quit because of AI.” The real fight is cram school vs. school-pace learning. Here is what the Chinese thread argued — and a direct ask for non-Chinese readers.
- How to Spot AI Writing in Chinese A V2EX thread on a Zhihu ban shows why detector scores often clash with human taste—and which Chinese AI tells, fake sincerity, and soft-promo patterns actually work.
- thisFeelingNeverGoesAway: the title comforts; the comments audit what’s left A ProgrammerHumor meme about the eternal coding feeling blew up. The top comment wasn’t “same”—it was “then his son discovers AI.” Read next to an earlier essay on accelerating vs replacing thinking: output rose; where did the aha go?
- Will AI Develop Self-Awareness? A V2EX thread shows why “I want to feel the world” is weak evidence—and why embodiment, refusal, and the definition problem matter more than picking a camp.
- Why AI suddenly got good at math: humans prune with taste, AI prunes with verifiers AI has been knocking down open math problems this year. A Latent Space essay offers a framework for predicting which problem falls next: difficulty depends on how many key hints a problem needs and how much work sits between them; humans search breadth-first and prune with taste, AI searches depth-first and prunes with verifiers, and failed attempts become the next round's hints. I walk through it, then add five thoughts of my own: verifiers set the boundary of AI capability, the failure list is a harness's most valuable asset, the human role shifts from digging to reframing, results may arrive without understanding, and the compute bill is real.
- The eternal complement: genius machines, boring work The first essay in OpenAI's "next economy" series argues that ideas and execution are complements. AI first makes execution cheap and taste scarce, but once AI produces good ideas on its own, execution may become scarcer than ever, and superintelligence may spend most of its time being the bureaucracy. I checked it against my own notes: the bottleneck is already visible inside companies, AI as bureaucracy is not AI imitating a bureaucracy, and depth domains move long before width domains.
- OpenAI now watermarks text: math barely takes it, edits erase it To comply with the EU AI Act, OpenAI has started adding textGrain watermarks to ChatGPT text. Its own numbers show detection is substantially lower for math, and swapping 25% of words drops detection from 92% to 17%. Using the technical report, this post explains why: a watermark can only borrow from the randomness in word choice, math has none to lend, and edits break the signal in runs. Then: who it actually catches, and what it cannot prove.
- AI panic: is our intelligence locked in? A game developer with 30 years of experience says everyone he asks feels they are getting dumber; on V2EX, people say programming is dying and even "hard work" no longer makes sense. The real difference is not whether you use AI, but whether you use it to accelerate your thinking or to replace it.
- How much AI-written code should you review? Go by risk A V2EX user asked: after AI changes your code, do you read the diff? Of 94 replies, fewer than one in ten review carefully, yet they got the most likes. Using code review research plus DORA and Veracode data, here is a risk-based review ratio and six things worth checking every time.
- Will pricier RAM push people to Linux? Where each OS goes with AI A V2EX user measured Windows using 4 to 6GB more RAM than Linux and asked whether rising memory prices will lift Linux adoption. Most of the 95 replies said no, and the data backs them. But pricey RAM is pushing OS vendors to slim down, and AI is rewriting what operating systems compete on. Here is how Windows, macOS, Linux, Googlebook, and HarmonyOS may evolve in the AI era.
- Small-town kids in the AI era: info is free, people aren't A V2EX parent asked how kids in a small county town can broaden their horizons. The 194 replies argue hard, but they describe four layers of the same thing: information, judgment, exposure, and hands-on work. AI has nearly closed the information gap, but not the gap in who you know. Drawing on CNNIC, PNAS, the World Bank, and Chetty, here is which layer small-town kids should work on now.
- What can programmers still do now that AI writes code? Two popular V2EX threads in one week: one user had Codex add ray tracing to their engine on the first try and asked if AI is replacing skilled programmers; another asked how many will still live off writing code in five years. 200+ replies and Stanford employment data say the squeeze is on entry-level jobs and on coding itself. Programmers can still do five things, and they were always most of the job.
- Do Codex quota resets cost you? A game-theory ledger A 195-reply V2EX fight: are quota resets a gift or a trap? It depends on whether they move your refill date. Keep the anchor and it is a gift; restart the window and it is a transfer — from people who pace themselves to people who burn fast.
- Why fan-translation crews may fear AI the least A Kazix interview piece: subtitle and manga fans don’t resist AI. When the goal is “more people can understand what we love,” automation of grunt work—and even the end of the crew—can be good news.
- Failure’s Worth, Through V2er Metaphors: Waterfall, Helicopter, Map A V2EX thread on Fields medalists warning about AI cracking math problems. The real gold is the metaphors—hiking vs helicopter, copying exam answers, fog-of-war maps, and inventing the crane.
- Distillation Through V2er Metaphors: Fishbowls, Kitchens, and Copying A V2EX thread asked if model distillation is like fishing from someone else’s bucket. The replies turned an abstract term into vivid metaphors—and a clean takeaway.
- When Average People Feel AI: Electricity for Knowledge Work, a Rounding Error Outside Nathan Lambert: the AI revolution may last a century, but daily life barely notices. Benefits are too indirect; politics looks like Engels’ pause. Deep adoption, not chat, marks the real shift.
- An Alien Mind: Alignment Unsolved—Don’t Scale Flat-Out OpenAI’s chief scientist: AI is a grown alien intellect; goal ≠ value alignment; CoT monitoring is weakening. No lab has solved alignment enough for max-speed scaling.
- ChatGPT Improves Answers; Critical Thinking Widens Ideas A 1,000-student RCT from Bocconi and OpenAI: AI raises polish; causal-reasoning training raises originality. Old rubrics miss half the signal.
- Will AI Shrink Total Jobs? Demand vs Productivity A V2EX thread argues jobs track demand, not productivity — and human desire forever outruns AI. The comments tear into effective demand, prices and deflation.
- What Self-Improving Agents Actually Improve: Warp vs OpenAI Warp’s two-skill feedback loop is shippable; OpenAI’s shared-memory incident is not proven RSI. Skills compound from human labels—not weight updates.
- DeepSeek Engram: conditional memory as a new sparsity axis DeepSeek Engram adds O(1) lookup for the conditional memory MoE never had. At equal parameter and compute budget, reasoning gains more than rote knowledge.
- Software Is Not Files: Finding Materials ≠ Having Knowledge vivo Xiao Bo's KDC part 2: representation is not knowledge. A top RAG hit on a stale refund policy gave a confident wrong answer — governance beats tuning.
- Nvidia's Risky Business: AI Funding Enters the Danger Zone Stratechery deep dive: from 1873 railroad bonds to hyperscaler debt and Nvidia's $500B GPU financing platform, each funding layer is riskier than the last.
- The Defense Window Is Closing: Daybreak and GPT-5.6-Cyber OpenAI answers AI-scale attacks with Daybreak Blue/Red tiers and GPT-5.6-Cyber, taking advanced cyber task completion from 1.5% to 95% and finding real CVEs.
- Green Dashboards, Worse Answers: Latency Is Correctness Flat error rates hide thinner RAG context and truncated agent loops. Why TTFT, tail latency and retrieval metrics matter for answer quality, not just speed.
- Before You Let Agents Run Loose: AI Context Architecture Context architecture is not the RAG pipeline — it's guardrails, scopes, trust scores and human-in-the-loop. Stack Overflow on infrastructure vs architecture.
- Wide + Deep: why a 4B model can punch up DeepSeek taught models to think deep. A Tsinghua team says deep isn't enough — you also need wide. How a 4B setup can stand next to a 671B.