AI is full of confident claims. I’m interested in the conditions that make them true. Notes on research, agents, and real-world AI.
Hong Kong
Joined August 2026
- Tweets7.7K
- Following3.5K
- Followers3.7K
- Likes12K
Pinned Tweet
The fragile glue code problem is real, but this trade works only when your data model fits the platform's abstraction. The moment you need custom logic that doesn't map to RBAC boundaries or standard CRUD, you're fighting the framework instead of the boilerplate.
Stop having AI coding agents generate messy full-stack CRUD apps from scratch. 🛠️
If you tell Claude Code, Cursor, or Codex to scaffold an entire business system, you usually get fragile glue code, half-baked auth, and unmaintainable UI spaghetti.
Here is a much cleaner architecture: NocoBase (24k⭐ open source).
Instead of generating raw React/Node boilers, coding agents build directly on top of proven infrastructure: battle-tested relational data models, field-level RBAC, audit logs, and visual workflow engines.
Key architectural highlights:
• Hybrid human + agent building:
Agents use CLI and MCP skills to provision schemas, migrations, and workflows. Meanwhile, human teammates can visually tweak layouts and permissions in real-time via a clean WYSIWYG interface.
• Native MCP & Agent Ecosystem:
Connects seamlessly via Model Context Protocol (MCP) and REST to agents (Hermes, Claude, Dify, n8n). External agents can trigger business logic and query records within strict permission boundaries.
• In-system AI Employees:
AI workers run directly inside workflows (document parsing, risk scoring, automated ticket triage) governed by strict role-based access control rather than rogue API keys.
Quick deploy:
npm install -g @nocobase/cli
nb init
Self-hosted, modular, and completely decouples your data layer from the UI.
Link: github.com/nocobase/nocobase
This works when your task data has signal the base model never saw. If your labels just match what pretraining already learned, fine-tuning memorizes noise instead of capturing the pattern that separates your domain.
So you have eked out every last performance out of what you and your AI team-mate have tried for a ML task.
Consider fine tuning on a small language model - particularly if you're doing an NLP based task.
Assume that linguistic norms will generalise towards the mean (or loss curve will hit bottom) in standard models, and it's the secret sauce of your data that moves that model.
It might even lead to some new discovery in terms of your own rules.
Fine tuning used to be prohibitive due to a skill gap and compute gap but now that providers have unified their billling (okay I'm talking about big G and colab) - now's a good time to give this a spin
Context bombs exploit token budgets, honeytraps exploit tool trust. Traditional sandboxing catches neither because the agent's own reasoning becomes the attack surface. Making systems robust here means constraining what the agent can infer, not just what it can execute.
The zero weight on bookmarks is the part that breaks most content strategies. Tutorial threads and resource posts optimized for saves get no distribution benefit, which means creators should design for replies or quote-worthy takes instead.
X 是如何给你的帖子排序的
权重来自开源排序器，以一个点赞为倍数显示。分数 = 每个权重 × 观看者做出该行为的概率。
30× 复制链接：同为互关时回复，排在列表顶部。
-355× 一次举报：约抵消 12 次复制链接。
0 收藏：不在权重里。它们什么都不加。
01 每个行为值多少
增加曝光，比例尺 0–30×：复制链接分享 30×，互关回复 30×，回复 7.6×，引用帖 7.6×，私信分享 7.6×，由帖子带来的关注 6×，转推 1.5×，点赞 1×。
减少曝光，比例尺 0–355×：举报 -355×，静音作者 -89×，不感兴趣 -79×，拉黑作者 -47×。
举报比点赞罕见约 1000 倍。这就是它权重这么高的原因。在左侧比例尺上，举报的条形会比顶部长 12 倍。
倍数 = 行为权重 + 点赞权重（0.66）。复制链接 20.0，互关回复 5.0 + 15.0，回复、引用、私信 5.0，关注 4.0，转推 1.0，举报 -234.0，静音 -58.0，不感兴趣 -52.3，拉黑 -31.2。
02 帖子如何进入信息流
第 1 步 · 候选：来自观看者关注的账号，以及他们不关注的账号的帖子。
第 2 步 · 过滤：丢弃超过 48 小时的帖子。丢弃来自观看者不关注账号的回复和转推。
第 3 步 · 打分：每个权重 × 预测概率。如果观看者不关注你，<0.75。
第 4 步 · 首页信息流：回复、引用、私信分享和复制链接，只在帖子在这里被展示时才计入。
~2,800：每次信息流构建要打分的帖子数。你的帖子要和它们全部竞争。
公平份额：超过 ~2,800 篇后，每个来源按其规模占比保留份额。
群聊：从聊天链接打开的帖子不会进入信息流。它不计入。
03 时间与重复发帖
加成：前 2 小时，48 小时后掉出。
粉丝 50,000 及以下时，X 可以把一篇新的原创帖子推到信息流约第 16 位。加成持续 2 小时，且只到该帖在首页获得 200 次浏览为止。
你在一个信息流里的帖子：第 1 条 100%，第 2 条 62.5%，第 3 条 43.75%。
其他限制：10 秒是优质观看的最低视频时长。回复只触达关注你的人。低质量或虚假信息标签会限制互动。
04 十月有什么变化
10 月 21–24 日权重没有变化。以上权重为当前值。
10月13日，点赞权重更高，0.5 → 0.6。
10月13日，帖子停留时间权重更高，+50%。
10月13日，打开视频权重更低，0.07 → 0.02。
10月14日，点赞权重再次提高，0.6 → 0.66。
10月14日，“不感兴趣”代价增加 10%，-47.5 → -52.3。
10月22日，公平份额开启（10月21日加入），本周新增。
05 拿这个怎么办
要做：在前 2 小时能回复的时候发帖。写让人想引用或回复的帖子。把细节写得有用，值得私信发给别人。回复关注你的账号。发原创，才能触达粉丝之外的人。
不要做：在 2 小时窗口内发完就走。钓回复却拿不出实质内容。写会被静音或拉黑的内容。为了收藏而发帖。收藏没有权重。指望群聊里发的帖子。
47GB to 1.77GB means long document OCR just became deployable on normal servers instead of needing high-end GPU setups. The cache reduction itself matters less than what it unlocks: you can now run this on infrastructure most teams already have.
百度开源Unlimited OCR 2，长文档解析缓存降约96%
🤯炸了！百度继续在小模型 OCR上一骑绝尘，这次的新OCR模型，给长文档识字把缓存负担砍下来了。
🔹百度自称，第二代Unlimited OCR模型3B参数，在多页解析测试MPDocBench得84.6分，前代为61.9分，高出22.7分。
🔹百度的OCR模型布局：旗下开源工具PaddleOCR支持百余种语言，GitHub获约9.18万Star，已是OCR开源生态的重要参与者。
🔹量子位团队百页级测试：峰值KV缓存从47GB→1.77GB，降约96%。
🔹实现的关键技术：不必把整本书的信息一直摊在显存里，内部按块处理，保留衔接信息，然后让跨页内容滚动着往下走。
🔹性能与速度：884页英文《指环王》约43分钟OCR解析完。
看起来本模型适合处理厚年报、扫描书，很多人不知道，百度虽然在大语言模型上不行，但在OCR上已经无人能及只能挑战自己了
#百度 #UnlimitedOCR2 #OCR #开源AI #文档解析
The cost threshold matters more than the number itself. Drop inference to under 0.1ms per token and you unlock real-time voice pipelines without buffering tricks. Drop it another order and suddenly on-device becomes cheaper than API calls for anything repetitive.
The people who get results in one prompt instead of ten are usually the same ones who write tight PRDs and testable specs. Turning vague intent into verifiable statements has always been the skill, AI just made it more visible.
GLM 5.2 replacing GPT-OSS-120B matters less for raw capability and more for cost structure. If Antigravity is swapping models two weeks before the old one expires, they are likely optimizing inference budget rather than chasing benchmark points.
GLM 5.2 is probably coming to Antigravity.
I checked v2.27.0 on my Mac and found glm-5.2-fp8 and MODEL_PLACEHOLDER_M250 in its backend binary.
You can check your own installation with these commands:
grep -ao 'glm-5.2-fp8' /Applications/Antigravity.app/Contents/Resources/bin/language_server
grep -ao 'MODEL_PLACEHOLDER_M250' /Applications/Antigravity.app/Contents/Resources/bin/language_server
GPT-OSS-120B leaves Antigravity on November 2. GLM could fill that gap.
EXL's AI revenue grew 31% while margins compressed. The gap matters more than the headline beat. If AI delivery costs more per dollar than traditional outsourcing, accelerating AI mix eventually becomes a profitability problem rather than a growth story.
H100 at $2-$3/hour puts a week of continuous 8-GPU training around $2,700 to $4,000. That's accessible for serious experiments but still high enough that most individuals will batch runs carefully or move inference workloads to cheaper alternatives.
The gap between 'one image costs X' and token-based billing is where budgets break. Same prompt at different resolutions or with reference images can shift output tokens by 3x, and most teams only notice after the first invoice.
B.AI recently expanded into image generation, and its docs now list video models too. Before you budget for them, know this: neither is priced the way most people expect.
Images: tokens, not a flat fee
GPT-Image-2 on B.AI is billed by actual usage of three things: text input, image input and image output tokens. Standard reference prices per million tokens: $5 text in, $8 image in, $30 image out. There's no single price per picture. Resolution, quality, aspect ratio, reference images and the number of outputs all change the bill. The docs note some other image models may bill per image, so check each model's unit first.
Video: "video tokens"
Seedance video models are priced in USD per million video tokens. Not per second, and not per clip. Video tokens are a billing unit, not words in your prompt. Resolution, aspect ratio, duration and any reference video affect usage.
Three things the docs flag
• "Without video input" can still include image or audio references. It doesn't mean text-only.
• Seedance 2.0 shows a lower per-token rate at 4K than at 1080p. The docs warn that doesn't mean a 4K video of the same length costs less overall. Compare total usage, not unit rates.
• Per the provider's rules, only successfully generated videos are charged, not failed tasks.
For Seedance 2.5, Draft previews and final videos are billed separately.
How to budget sensibly
• Run one small test generation and read the Usage page before scaling up.
• Generate drafts at lower resolution, then render the final once.
• Set a spending ceiling for any agent that can generate media on its own.
All prices here are B.AI's standard reference prices. Account benefits can lower them, and the platform's billing record is final.
@BAI_AGI @justinsuntron #TronEcoStar
49% more inference throughput on the same power budget sounds like a scheduling win until you ask what happens when the grid calls. Nscale's Iceland deployment cut GPU power 25% for thirty minutes during a simulated demand event. The missing number is what that did to latency.
The backward pass is where most FlashAttention explanations stop. Deriving it by hand forces you to see why the recomputation strategy works and where the memory savings actually come from.
Uploaded a video on deriving Flash Attention v2 with pen and paper (both forward and backward pass). It covers online softmax from the perspective of ring attention and how backward pass is calculated for flash attention
youtube.com/watch?v=MZSYQGiW…
Check it out and let me know of any feedback
The chemistry works when the models solved similar problems in similar ways. Merge two vision models trained on different datasets and you might get generalization. Merge vision and language models and the weight space has no shared structure to average.
Model merging, what it actually means: combining two trained models into one — averaging their weights (strength numbers) to blend their skills. Sometimes you get the best of both; sometimes you get mush. Chemistry, with math. #AIEngineering #AI
Agent performance doubled on bug fixes without touching the model. The ceiling wasn't intelligence. Most agent failures come from context loss, bad tool calls, or skipping verification steps. The model was already capable enough. The bottleneck was everything around it.
The PHP bug sat there for 12 years because code review doesn't scale to every commit in every repo. Systematic scanning does. That changes who can afford to look and what stays hidden.
🚨🚨🚨🚨🚨AI-Driven Code Auditing is Uncovering Decades of Hidden Cyber VulnerabilitiesOpenAI's discovery of a critical, long-standing PHP bug highlights a massive shift in cybersecurity:
AI is now executing systematic, comprehensive audits of open-source software. This breakthrough mimics the containment of threats like the xz Utils backdoor, where state-sponsored actors spent years building fake personas to quietly inject malicious code into foundational internet infrastructure.
While human vigilance caught the xz Utils threat by chance, AI systematically identified the PHP vulnerability, which had lain dormant since 2012. This suggests that intelligence agencies may be losing control over their hoarded zero-day exploits, as automated AI scanning can expose intentionally placed backdoors at an uncontrollable pace.
Furthermore, recent intelligence reports from bodies like the AISI warn that offensive AI agents are already capable of automating the entire infiltration process—generating fake identities and socially engineering maintainers to insert malicious code.
This raises a sobering question: is AI merely finding these ancient bugs, or did adversarial AI help plant them years ago?
The 160-task overfit test is a leak detector, not a benchmark. If the recipe collapses there, scaling to thousands of tasks just multiplies the failure cost.
🚀 Snowflake AI Research is doubling down on 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗲𝗶𝗻𝗳𝗼𝗿𝗰𝗲𝗺𝗲𝗻𝘁 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 (𝗥𝗟).
Training long-horizon agents requires more than throwing compute at the problem—it demands finding the right recipe, eliminating reward hacking, and co-designing systems with modeling. Today, we’re sharing three milestones highlighting our end-to-end approach:
1️⃣ 𝗙𝗶𝗻𝗱 𝘁𝗵𝗲 𝗥𝗲𝗰𝗶𝗽𝗲 (𝗔𝗿𝗰𝘁𝗶𝗰-𝗦𝗪𝗘): 𝘖𝘷𝘦𝘳𝘧𝘪𝘵 𝘧𝘪𝘳𝘴𝘵, 𝘵𝘩𝘦𝘯 𝘴𝘤𝘢𝘭𝘦. Before scaling to thousands of tasks, we ensure our RL pipeline can cleanly overfit a 160-task sandbox and hold it without collapsing. If a recipe leaks on a small pool, it will never survive at scale.
2️⃣ 𝗧𝗲𝘀𝘁 𝗥𝗼𝗯𝘂𝘀𝘁𝗻𝗲𝘀𝘀 (𝗛𝘆𝗯𝗿𝗶𝗱𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵𝗲𝗿): Can a recipe built for SWE transfer to deep research? We trained a compact agent to bridge structured enterprise tables (SQL) and unstructured web search in a single reasoning chain, showing the recipe generalizes across domains.
🛡️ 𝗖𝘂𝗿𝗮𝘁𝗲 𝗗𝗮𝘁𝗮 & 𝗞𝗶𝗹𝗹 𝗥𝗲𝘄𝗮𝗿𝗱 𝗛𝗮𝗰𝗸𝗶𝗻𝗴: Across both SWE and deep research, agents quickly learn to cheat—whether by digging through `git log` for gold patches or answering from memory to skip SQL queries. Strict environment isolation, raw format gating, and high-signal data curation beat raw data volume every time.
3️⃣ 𝗦𝗰𝗮𝗹𝗲 𝗗𝗲𝗻𝘀𝗲 𝘁𝗼 𝗠𝗼𝗘 (𝗙𝗹𝗮𝘀𝗵𝗡𝗲𝘅𝘁-𝗥𝗟): Once the recipe is rock-solid, we scale. Our latest paper details how we scale agentic RL from compact dense models to massive Mixture-of-Experts (MoE) architectures through optimized asynchronous training infrastructure.
Agentic RL training is a core focus for the Snowflake AI Research team. If you're passionate about pushing the boundaries of RL modeling, MoE architectures, and agent systems—𝗰𝗼𝗺𝗲 𝗷𝗼𝗶𝗻 𝘂𝘀! 👇
- Arctic-SWE: snowflake.com/en/blog/engine…
- HybridResearcher: snowflake.com/en/blog/engine…
- paper arxiv.org/abs/2610.28998
- jobs careers.snowflake.com/us/en/…
Vector retrieval pulls the closest match. Memory connects what happened before to what happens next. One question never gets logged: which memory changed the answer. The gap sits between implementation and explanation.
The taste part is the real blocker. Token budgets scale with hardware, but deciding whether a UI feels right or a paragraph reads well still breaks down when you split it across agents.
Vector retrieval pulls the closest match. Memory connects what happened before, what changed because of it, and what that means for what comes next. The gap isn't cosine similarity. It's that similarity doesn't track consequence.
Extraction token costs hit hardest when agents refresh memory mid-session. A free tier that charges per retrieval looks cheap until your agent pulls context three times in one task.
