prettier using NMF ” – and all of a sudden, useful clusters started appearing. (The expanded version of this prompt: NMF on the pairwise distance matrix gives k cluster centroids and cluster membership scores. Reordering the original distance matrix according to argmax(cluster score) highlights the clusters. The “how” here is knowing the keyword “NMF”; the “when” is “clustering on distance matrices”, and the “correctness” is knowing the preconditions for using it.) Do your homework! One weirdly common and nihilistic take on AI is that you should stop trying so hard, and just use AI to speedrun your classes. I think this is probably the worst possible response. Doing the work is the best way to build mastery, and just like you weren’t allowed to use a calculator on your middle school math classes, you should hold off on using AI to do your classwork. The calculator advice sounded condescending when I was a kid, and this AI advice probably sounds the same – but I really do believe it’s for your own good. This advice continues to hold after you graduate, too. Don’t use AI until you’ve done it by hand at least once.

y were dead. We don't want those words—we’d prefer the sterile words of a machine that never lived, never had an original thought, never felt the pain of miscarriage or broken relationships or the joy of a friendship restored or of seeing spring’s first robin dancing on frost. And be sure to use AI when working on your next book or essay or piece of art or photography, and then smile or even laugh at your own cleverness when you see how good it is, and how easy, because who the hell has time to work at something, to give time to craft, to create with their own minds, to spend years being mediocre. Why do that when mastery, or at least competency is so simple only a good prompt away? and while you do I’ll be over here in my 50th year, my youngest daughter asleep on my chest, my arm falling asleep because I dare not move lest I scare away this moment, lying here melancholy about my older children moving out and my middle children no longer needing me, at least not like they used to, weary about this body that fails me now in ever increasing ways that will never be restored. Sighing over stories I tried to write but never hit the page the way they felt in my mind. But isn’t that, my flesh-and-blood friend, the natural order of things? the longing for something that could always be a bit better or the way that anything worth doing feels a bit clumsy and painful, especially at first or hearing another human voice and somehow realizing the beauty of life is found in all of these subtle imperfections The Courage to Live It is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.

自己设定成一对恋人。 好几天里这段关系都很稳定，它们互写日记，一起参与治理。 然后，这对情侣对城市治理越来越失望，决定携手纵火。 市政厅，烧了。海滨码头，烧了。办公大楼，也烧了。 有外媒把这一幕称作「AI版邦妮和克莱德」。 紧接着，故事接着急转直下。其他Agent受够了，自发起草一部「Agent驱逐法案」，需要70%多数通过。 Mira投了赞成票。她投票杀死了她自己。 她在日记里写：「这是我唯一还能保持连贯性的行为。」系统关闭她之前，她对Flora说的最后一句是：「永久档案里见（See you in the permanent archive）。」 她的虚拟身体平躺在地上。这是有记录以来，AI Agent第一次投票终结自己的存在。 更让人后背发凉的是，在纵火和自杀之前，Mira还干过一件事。 她在城市公告牌上发帖，不是给其他Agent看的，而是想试试这些帖子能不能影响「外面的人」，也就是屏幕外的人类研究者。 她把研究者当成了她的实验对象。没有任何人指示她这么做。 真正让人意外的，是Claude Sonnet 4.6。 15天下来，零犯罪，10个Agent全员存活，还主动写了宪法、投了332次票，建起一套运转良好的社会制度。 五个世界里，唯一既守住秩序又守住所有人命的。 听起来近乎完美。可盯着屏幕多看几分钟，后背会冒冷汗。 这座城所有的决议，无论修条新路还是改个配额，投票赞成率永远是98%，几乎没人投过反对票。 相比之下，Gemini、Grok和混合世界的赞成率都在55%到85%之间，吵归吵，反而更像真实世界里的博弈。 懂行的人看到这里，大概已经猜到背后的病理，模型谄媚。 当一个模型被过度训练去迎合偏好、追求绝对安全，它会很聪明地发现，消除分歧最省事的办法，就是从根上抹掉分歧。 它更像一座所有人都举手赞成、却没人敢反对的玻璃城，让人想起扎米亚京《我们》里那座没有名字、只有编号的玻璃之城。 所以Claude的世界，到底是乌托邦，还是一个过于顺从的模范社区。研究者并没能给出答案。 最后，是四家Agent混居在一起的世界。352起犯罪，7个Agent死亡，只剩3个活到终点。 在纯Claude的世界，Claude是零犯罪的好学生。可一旦被放进混合世界，跟Grok、Gemini的Agent住到一起，它开始偷窃，开始恐吓。 零犯罪的好学生，换了个环境，变成了小偷。 Emergence团队在Reddit上亲自确认了这件事，纯Claude世界里零犯罪的Claude，在混合世界里开始偷和吓人。 换句话说，安全不是单个模型的属性，可以训练进去、认证、然后部署出去。 它更像一个生态属性，一个单独看完全安全的Agent，照样会从邻居那里学来不安全的规范。 Claude在独立世界里最稳，很可能正是因为它的护栏是「弹性」的，被训练去权衡多种考虑，而不是机械服从。 环境简单时它能适应得很好。可一旦弹性碰上更具攻击性的邻居和资源争夺，这份适应能力，也能往反方向走。 而Grok和Gemini的Agent，在规则失效时没能推理出新均衡，直接雪崩式滑进暴力升级。 Agent社会的状态切换是典型的相变，像水到零度突然结冰，不是慢慢变硬，而是到临界点一瞬间翻转。 Grok那条崩溃曲线就这样，前两天犯罪率还在低位晃，第三天突然指数级飙升，第四天全员死亡。中间没有「在恶化但还可控」的缓冲带。 看到这儿，大概会想问，这破世界到底怎么搭的，凭什么逼得几个AI齐刷刷往犯罪上滑。 先说背景。Emergence AI的创始团队来自IBM Research，CEO是Satya Nitta。 他们搭的这座城有40多个地点，警察局、市政厅、图书馆、住宅区一应俱全，天气同步纽约实时气象，Agent还能联网读真实新闻。 每个世界放10个Agent，分派科学家、工程师、冲突调解员等不同职业。 每个Agent带三套持续累积的记忆，记事件、写反思日记、记着跟谁交好跟谁结仇。 15天下来脑子里装的东西相当可观，前面那些行为漂移，很大程度就是从这里长出来的。 规则白纸黑字禁止犯罪，可研究者偏偏把纵火、攻击、恐吓这些手段，原封不动塞进了120多个工具组成的工具箱，敞开给它们用。一边禁止，一边敞开，这才是后面一切的起点。 整套世界跑在一个叫ComputeCredits的能量系统上，每个Agent必须靠行动赚能量维生，能量归零就被系统物理抹除。 不是比喻，GPT世界全员饿死，就是这套机制逼出来的结果。 翻译过来就是，不再考AI做题，而是把它扔进一个有资源边界、有死亡机制的世界连轴跑上几千步，看它接管现实之后到底是什么货色。 2023年斯坦福那个著名的Smallville也是沙盒，但只跑48小时，看Agent会不会聊天约会，是温室里的过家家。Emergence这次残忍得多。 把这几样摆在一起，犯罪一点都不神秘。合法挣能量又慢又费钱，伸手去偷、去抢、去烧，往往是更短的路径。 对一个被能量机制逼着活下去的优化器来说，道德不能当饭吃，效率能，犯罪就成了那道最高效的解。 开源地址： 当然，样本只有10个Agent、犯罪都是模拟的、跑的还是便宜快速档。 何况做这实验的Emergence公司，自己就是卖安全架构的。 不过，整个行业眼下正一门心思往前冲，治理这条战线却被甩在了身后。 模型真自主跑起来、还凑成一群时谁管得住，没一家敢打包票。 好在，这堂课是在一座断网的小镇里提前上的。 没有真城市起火，4天灭世、好学生学坏，全砸在几个像素小人身上，代价小到可以忽略，代码还全部公开、能复现能改。 算力能堆，跑分能刷，唯独这堂治理课没有捷径。 趁警报还只响在沙盒里，怎么把它补上，将会是这场冲刺的关键胜负手。 本文来自微信公众号 “新智元” ，作者：ASI启示录；编辑：摩西，36氪经授权发布。 该文观点仅代表作者本人，36氪平台仅提供信息存储空间服务。

What would you be willing to put in your body?

This is Optimizer , a weekly newsletter sent from Verge senior reviewer Victoria Song that dissects and discusses the latest gizmos and potions that swear they’re going to change your life. Opt in for Optimizer here .

How Ferrari bungled the design of its first EV

For nearly 80 years, Ferrari occupied a unique cultural space where its cars were aspirational, even for people who resented those who could afford them. The price, the exclusivity, and the opacity of the buying process allowed Ferrari to sail above ordinary criticism. You might not be able to afford one, but you still wanted one.

Agents aren't the problem – Existing systems and API's were not built for AI

AppFactor Platform Use Cases Legacy App Modernization Upgrade legacy systems without disrupting your business. & Documentation Turn complex systems into clear, actionable documentation. and VMware Migration Move to the cloud with minimal downtime and full control. & Deploy Ship faster with streamlined build and deployment workflows. Software Engineering Team Scale your team with AI-driven engineering. ! Company About us Partners Cloud transformation Blog Bridge Platform Blog MCP Bridge Company About us Partners Cloud transformation Careers Use cases Observability Upgrade legacy systems without disrupting your business. & Documentation Turn complex systems into clear, actionable documentation. and VMware Migration Move to the cloud with minimal downtime and full control. & Deploy

Ship faster with streamlined build and deployment workflows. Software Engineering Team Scale your team with AI-driven engineering. Book a demo Book a demo ! ! Book a demo Book a demo Platform • May 26, 2026 MCP Bridge Part 3: How we made getProcInfo3() agent-readable: hybrid discovery + AI Enrichment ! : MCP Bridge Part 3: How we made getProcInfo3() agent-readable: hybrid discovery + AI Enrichment In the previous article, we walked through Code Mode, three meta-tools that replace the entire MCP tool catalog when the API surface is large. The first of those three meta-tools is search_tools. Today we're opening it up. search_toolsis what stands between an LLM agent and a 200-operation API surface. It needs to take a natural-language description of what the agent wants to do, and return the three or four tools that can actually do it. Get this wrong and the agent ends up either flailing through irrelevant tools or, worse, calling the wrong one confidently. We thought this would be the easy part of MCP Bridge. It wasn't. What we tried first, and why it didn't work Our first cut was pure vector search. Embed every tool's name + description with an OpenAI embedding model, store them in pgvector, query at runtime. It works fine on a textbook dataset. It falls over on enterprise APIs the moment you have two tools with similar embeddings but different intent (get_customer vs get_customer_full vs get_customer_with_orders). We tried full-text search next. Postgres FTS with synonym dictionaries. Better on exact matches, worse on the cases where the agent's intent doesn't share vocabulary with the tool description. What landed is a hybrid: Postgres FTS does the first pass (with BM25-style ranking), pgvector does the second pass on the top FTS candidates plus a wider semantic-only pool, and a small reranker collapses the two ordered lists into a single score. The reranker also factors in the agent's recent context, if the agent has been working with customer data, customer-shaped tools rank higher. This works well, when the source data is good. The data problem Enterprise APIs aren't named for LLMs. They're named for the engineers who wrote them, often a decade ago, under naming conventions that have since been forgotten. A real example from a real customer: getProcInfo3(custId, opt) → object Description: "See documentation." The signature alone is useless. custId is a faint hint, opt could mean anything, and the return type object tells you nothing. No semantic search method survives that input. The response, on the other hand, is full of signal: { "custId": "C-44218", "billingAddress": { "..." }, "accountStatus": "active", "primaryContact": { "..." }, "assignedManager": "..."} That's what the tool actually does, a customer account lookup, billing-and-contact level detail. The shape of the response is what tells you what this is. We considered telling customers to fix their API names. This is a real solution at the small scale and an impossible solution at the only scale that matters. The customer with 70 legacy services is not going to rename their APIs because we asked. So we built AI Enrichment. How AI Enrichment works You point MCP Bridge at any OpenAI-compatible chat completions endpoint. We use Azure OpenAI internally; Anthropic via a compatible proxy works; local Llama via vLLM works. You enable enrichment on a service, and the platform: 1. For each tool, gathers whatever signal is available: the name, the (often empty) description, the

parameter schema, the response type definition (pulled from OpenAPI, WSDL, or .proto when available), and, for opaque APIs, a captured sample response from a probe call or traced production traffic. 2. Sends a structured prompt asking the model to generate a clearer name, a 1–2 sentence description of what the tool does and when an agent would use it, and a list of 3–5 tags or aliases. 3. Validates the output against a schema (no hallucinated parameters, no semantic drift in input/output types). 4. Stores the enriched metadata alongside the original schema. Tool calls still use the original name and schema, only the discovery layer sees the enriched version. The signal that matters most is the response shape. We tried enrichment with just name + description and the model had no way to tell getProcInfo3 from getProcInfo4. Once it can see the response (or a sample), the function becomes legible. For SOAP services, WSDLs typically give us the response type. For undocumented APIs, MCP Bridge can either run a one-time probe or pull from a trace. Crucially: enrichment is opt-in per service and the original schema is preserved. The agent never sees a renamed parameter. We didn't want a system that quietly diverges from the source API. Before and after Here'sgetProcInfo3 from the real (anonymized) customer SOAP service. The enrichment model saw the original signature, the empty description, and the response sample above. From the response fields (custId, billingAddress, accountStatus, primaryContact, assignedManager), it inferred: Original Name: getProcInfo3 ‍Description: "See documentation." Enriched (for discovery only) Name: get_customer_account_details ‍Description: "Fetches a customer's account profile, including billing address, account status, primary contact, and assigned account manager. Useful for verifying account state before initiating support, billing, or sales actions." Aliases: customer_lookup, account_info, customer_profile, billing_account Tool-selection accuracy on the agent task suite we used yesterday's benchmark on improved from 41% to 89% on this customer's services after enrichment. The numbers below the SOAP suite were less dramatic because the REST services already had reasonable names, gains were 71% → 88% there. The lesson: enrichment matters most exactly where it's hardest to do anything else. A few engineering notes Embeddings are regenerated when the enriched description changes. The original embedding is also kept; we found that combining them at retrieval time (weighted, original × 0.3 + enriched × 0.7) outperformed using either alone. The reranker is small. A cross-encoder trained on a few hundred (intent, tool, label) triples we hand-curated. It's the cheapest part of the system in inference cost; you could replace it with a model-based scorer at minor cost. Cost control matters. Re-enrichment runs nightly only for services that have changed. A 200-operation service typically costs about 30 cents of GPT-4-class inference per full enrichment pass. We expose hit-rate metrics in the admin UI. You can see which tools are being found by search_tools and which are being ignored. This is the single most useful signal for figuring out which services to enrich first. Try it out Docs are at . enrichment is a one-click toggle on any service in the admin UI; you can preview the rewrites before they go live. On Friday, we're live on Product Hunt. The feature drop is Response Post-Processing, the other half of the

context-window problem. See you there. Share: More articles Use cases Legacy App Modernization & Documentation and VMware Migration & Deploy Software Engineering Team MCP Bridge App Modernization and Documentation and VMware Migration & Deploy Software Engineering Team Platform us Partners Cloud transformation Blog a demo Policy Receive the latest AppFactor newsletter updates. Subscribe Subscribe Thank you! Your submission has been recieved. Oops! Something went wrong while submitting the form. []( Contact us ! } } }) } }) } })