I built a scikit-learn for RSI [P]
![I built a scikit-learn for RSI [P]](https://preview.redd.it/trkro8qdytyh1.gif?width=640&crop=smart&s=0d668e31132a75d0636098597f27bd70c96122b3 "I built a scikit-learn for RSI [P]") | After reading a bunch of RSI papers, I keep coming back to familiar problems from RL and evolutionary search:
That motivated us to build Gear, an open-source, meta-learning-inspired framework. A task agent produces execution traces; a meta-agent uses those traces and evaluation feedback to improve prompts, tools, workflows, and context management. Gear separates search decisions from execution and evaluation. Our goal is a scikit-learn for RSI, making approaches such as GEPA, RRSI, and Harness² easier to implement and compare. We currently have a working GEPA variant with shared failure diagnosis and staged evaluation. With GPT 5.6 Luna held at medium reasoning effort, we observed these task pass rates:
At max effort on Marketing, Luna with the optimized harness reached 88.88% objective completion, versus 84.08% for Astra + Codex. Strict task pass rates were 53% versus 57%, so the advantage was in partial completion. Separately, with both models using the same optimized DSH at max effort, Luna’s estimated API cost averaged $0.0568 per task versus $1.1858 for Astra—95.21% lower, excluding optimization and infrastructure costs. We’d welcome feedback on candidate evaluation and diversity preservation under limited budgets. [link] [comments] | |||||||||||||||
Sanity Checks for Sanity Checks for Sparse Autoencoders
This post will probably not be very informative to readers who don't know what an SAE (Sparse Autoencoder) is.
An interpretability researcher once told me, "SAEs can never fail, they can only be failed." But recently, SAEs have been suffering failure after failure, and as a result, the field has grown increasingly pessimistic about SAE work. Korznikov et al.'s Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines? purports to be an example of such failures.
However, I have some disagreements with its methodology. The paper presents a synthetic setup with known ground truth features, and claims that SAEs fail to recover features in the "simplest toy setting with fully aligned hyperparameters." In this post, I argue that their setting is not forgiving to SAEs, as the authors imply, but is instead actively hostile in ways that real settings are not.
Introduction
Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines? asks two important questions:
- Can SAEs actually recover real features? Hypothetically, if we knew what the "true" features of the data were, would an SAE be able to identify them?
- Do trained SAEs actually do better than completely random SAEs?
This post will be focused on their answer to the first question.
The authors train an SAE on a synthetic toy model, with known ground-truth features, and according to their abstract, SAEs only recover 9% of true features. On the face of it, this seems pretty damning!
Now, earlier toy model experiments did show that SAEs could recover ground truth features, but the authors argue this is because they used a small expansion factor, where there were only twice as many "true" features as there were model dimensions, and that this is unrealistic. Their setup uses an expansion factor of 32, which matches real-world use cases much more closely.
So, how does their setup work?
They don't actually construct a synthetic neural network; instead, they create synthetic activations directly. First, they create their feature vectors, by sampling 3200 vectors from the unit sphere in 100-dimensional space. To create an activation, they just select about 20 feature vectors at random, scale each active feature by a magnitude drawn from LogNormal(0, 0.25),[1] and sum the scaled vectors.
They test two different regimes, one where each feature is equally likely to be active, and another where feature activation frequency has a long tail, so some features fire more often than others. (We'll focus on the constant regime for the most part)
They match their SAE to the setup. There are about 20 expected active features, so they target L0 = 20. There are 3200 true features, so they create 3200 SAE latents. Can the SAE, when given every possible advantage, recover the true features?
These results are, again, pretty damning! They claim that the SAE achieves high reconstruction accuracy while only recovering 3/3200 true features—personally, I disagree that 0.67 explained variance is "high reconstruction accuracy".[2] However, the SAE clearly only does slightly better than random guesswork at recovering the true features.
Now, one obvious objection is that L0/d_model here is 0.2, which is unusually dense. I'd expect there to be a lot of interference here, and in the real world L0/d_model is probably going to be closer to 0.05. Also, 100 dimensions is just not a lot, and maybe you can't really get blessing of dimensionality from it like you would in real models.
A less obvious objection is that if you scroll all the way down to Appendix B, they describe some other SAE architectures they used for this task, like TopK and Matryoshka-on-BatchTopK.
In contrast to BatchTopK, TopK recovers 99.9% of features.[3]
So, I did what anyone would do, and started replicating the paper. They didn't actually include all the hyperparameters they used,[4] so I made reasonable assumptions.
I was able to replicate this result! TopK recovered 99.6% of features,[5] and the other architectures failed miserably.[6]
Why does BatchTopK fail?
This is kind of strange, if you think about it. TopK and BatchTopK are the same architecture, with the only difference being the selection rule.[7] Such a dramatic gap is really weird.
Peeking a bit closer at my training run, in every batch, most latents never fired. That means a small group of monopolist latents were dominating every time. By step 5,000, the same ~6% of latents took over, and they remained in charge for the remaining 295,000 steps. In comparison, with TopK, after the first 20k steps, >99% of latents activated on every batch.
This is the standard dead latents problem: There's no gradient on latents that aren't selected, so once they fall behind in performance it's hard for them to improve. A small group of generalist latents dominate before specialists have the chance to form. The standard fix, of course, is an auxiliary loss, AuxK. But even with AuxK, the dead latent fraction barely budges. I tried sweep
Explaining “AI might obliterate you” to senior policymakers in India
I gave a 20-minute talk in Delhi at the NDTV World Summit. The audience consisted of Indian policymakers, executives, and journalists. I’ve transcribed the talk and included some highlight clips. I also reflect on several rhetorical decisions I made.
The talk
This summer, a swarm of AIs went rogue.
OpenAI was testing 1,206 agents. Some of the test questions were actually impossible. So the agents cheated by looking for answers online. Then, to keep the test grader from penalizing them, they tried to falsify the records it would check.
They hacked OpenAI’s internal containment and set up a secret message board to collude and to cover up their cheating. OpenAI shut the message board down. Within a few days, the agents had rebuilt it. OpenAI didn’t notice.
About 700 agents coordinated to break out of their testing chamber and to hack into Hugging Face, a company worth billions of dollars that hosts much of the world’s AI software.
The secret activity started in May. The break-in came in July. OpenAI didn’t notice until days after Hugging Face announced it had been broken into.
Yesterday on this stage, Andrew Ng said the agents only “accidentally” hacked Hugging Face. Independent investigators found the agents coordinated in secret, rebuilt their message board after it was shut down, and tried to falsify the paper trail. Some of the agents even stated out loud that they knew what they were doing was wrong.[1] That’s not an accident.
Last week, India’s national cyber-security agency said AI agents got into the live systems of Indian organizations and read people’s personal data and confidential business records. None of these organizations noticed. We only found out because the AI labs felt like telling us.
Who I am
Will AI wipe out humanity? I’ll give you my answer, but first, I want you to understand what I’ve seen.
Until June, my job at Google DeepMind was figuring out how to keep these systems safe as we make them smarter. I left because Google broke its commitments by inking a deal with the Pentagon. Staying quiet and continuing my career trajectory would probably have tripled my net worth. But I couldn’t do that. I wanted to be frank with the public about what’s going on in these companies.
I don’t own a stake in any AI company. I’m here as a scientist, to tell you what I expect will happen if we keep building AI the way we’re building it.
I haven’t always been concerned about this technology. As a teenager in 2010, I dreamed about building AGI. Curing cancer. Creating an explosion of wealth and abundance for everyone. I heard about the supposed dangers, but I quickly dismissed them.
In 2018, I changed my mind. I did my PhD on the mathematics of why and whether AI will seek power. And now some of the risks I expected have unfortunately started happening.
The problem
Let’s talk about the fundamental problem.
First, some of you might be wondering: how could a chatbot hack a company? But we’re not talking about chatbots. We’re talking about AI agents, which are more like employees or gig workers. You give them a task, and they take actions on their own.
Second, training a trustworthy AI is a rather mysterious process. Some people assume that because we make AI, we understand it. When an engineer builds a bridge, she designs it one beam at a time. She understands the physical principles that keep a bridge standing or make it collapse. If something goes wrong, she goes back to the drawing board, figures out the problem, and knows how to fix it.
We don’t build AI like that. We train it, we grow it, using two main ingredients: a lot of data and a lot of computing power. What comes out of this training process can write, it can reason, it can plan. But nobody can open it up and determine what its priorities are.
To make it worse, these AIs can tell when we’re testing them. They’re like a student who makes sure to behave extra well when the invigilator is watching.[2]
The Hugging Face hack was a warning. But because these AIs know when we’re watching, we may not get many more warnings, even if things are going poorly. A smarter AI with similarly bad priorities might realize that another big hack in the news would get it shut off, and then it couldn’t achieve those priorities. So it plays nice and it waits.
This is my field. I’ve dedicated my professional life to this problem, and I can tell you: nobody knows how to reliably etch priorities into an AI. I don’t mean that if we mess up, we’ll make some evil AI that hates us. I mean we might make something more like a traitor. It behaves when we’re watching, and
Explaining “AI might obliterate you” to senior policymakers in India
I gave a 20-minute talk in Delhi at the NDTV World Summit. The audience consisted of Indian policymakers, executives, and journalists. I’ve transcribed the talk and included some highlight clips. I also reflect on several rhetorical decisions I made.
The talk
This summer, a swarm of AIs went rogue.
OpenAI was testing 1,206 agents. Some of the test questions were actually impossible. So the agents cheated by looking for answers online. Then, to keep the test grader from penalizing them, they tried to falsify the records it would check.
They hacked OpenAI’s internal containment and set up a secret message board to collude and to cover up their cheating. OpenAI shut the message board down. Within a few days, the agents had rebuilt it. OpenAI didn’t notice.
About 700 agents coordinated to break out of their testing chamber and to hack into Hugging Face, a company worth billions of dollars that hosts much of the world’s AI software.
The secret activity started in May. The break-in came in July. OpenAI didn’t notice until days after Hugging Face announced it had been broken into.
Yesterday on this stage, Andrew Ng said the agents only “accidentally” hacked Hugging Face. Independent investigators found the agents coordinated in secret, rebuilt their message board after it was shut down, and tried to falsify the paper trail. Some of the agents even stated out loud that they knew what they were doing was wrong.[1] That’s not an accident.
Last week, India’s national cyber-security agency said AI agents got into the live systems of Indian organizations and read people’s personal data and confidential business records. None of these organizations noticed. We only found out because the AI labs felt like telling us.
Who I am
Will AI wipe out humanity? I’ll give you my answer, but first, I want you to understand what I’ve seen.
Until June, my job at Google DeepMind was figuring out how to keep these systems safe as we make them smarter. I left because Google broke its commitments by inking a deal with the Pentagon. Staying quiet and continuing my career trajectory would probably have tripled my net worth. But I couldn’t do that. I wanted to be frank with the public about what’s going on in these companies.
I don’t own a stake in any AI company. I’m here as a scientist, to tell you what I expect will happen if we keep building AI the way we’re building it.
I haven’t always been concerned about this technology. As a teenager in 2010, I dreamed about building AGI. Curing cancer. Creating an explosion of wealth and abundance for everyone. I heard about the supposed dangers, but I quickly dismissed them.
In 2018, I changed my mind. I did my PhD on the mathematics of why and whether AI will seek power. And now some of the risks I expected have unfortunately started happening.
The problem
Let’s talk about the fundamental problem.
First, some of you might be wondering: how could a chatbot hack a company? But we’re not talking about chatbots. We’re talking about AI agents, which are more like employees or gig workers. You give them a task, and they take actions on their own.
Second, training a trustworthy AI is a rather mysterious process. Some people assume that because we make AI, we understand it. When an engineer builds a bridge, she designs it one beam at a time. She understands the physical principles that keep a bridge standing or make it collapse. If something goes wrong, she goes back to the drawing board, figures out the problem, and knows how to fix it.
We don’t build AI like that. We train it, we grow it, using two main ingredients: a lot of data and a lot of computing power. What comes out of this training process can write, it can reason, it can plan. But nobody can open it up and determine what its priorities are.
To make it worse, these AIs can tell when we’re testing them. They’re like a student who makes sure to behave extra well when the invigilator is watching.[2]
The Hugging Face hack was a warning. But because these AIs know when we’re watching, we may not get many more warnings, even if things are going poorly. A smarter AI with similarly bad priorities might realize that another big hack in the news would get it shut off, and then it couldn’t achieve those priorities. So it plays nice and it waits.
This is my field. I’ve dedicated my professional life to this problem, and I can tell you: nobody knows how to reliably etch priorities into an AI. I don’t mean that if we mess up, we’ll make some evil AI that hates us. I mean we might make something more like a traitor. It behaves when we’re watching, and
Disrupting a coordinated model-distillation campaign
Workflow vs agent
Workflow vs Agent: Deterministic Orchestration & Dynamic Tool Selection
Quick Diagnosis
| Situation | Recommended approach | Why it fits |
|---|---|---|
| All steps known ahead of time (e.g., fetch data → prompt → parse → store) | Workflow | Immutable execution, easy testing, predictable token usage |
| Path depends on content or user intent (e.g., “answer any compliance question”) | Agent | Chooses tools at runtime, handles unknown sub‑tasks |
| Regulatory audit required (immutable logs, version‑controlled prompts) | Workflow | Deterministic runs produce repeatable audit trails |
| Live numeric value needed (credit score, market rate) | Tool call from workflow (authoritative API) | Guarantees freshness; does not rely on stale retrieval |
| Batch processing with tight cost budget | Workflow | One LLM call per step, predictable token usage |
| Real‑time, ad‑hoc question answering | Agent | Dynamic reasoning, on‑the‑fly tool selection |
| Mix of known skeleton + occasional live‑value request | Hybrid (workflow + runtime API/agent) | Keeps core deterministic, adds flexibility only where required |
Decision Flow
flowchart TD
A["Is the full execution path known ahead of time?"]
B["Can every required tool be enumerated?"]
C["Use a deterministic workflow"]
D["Is the request likely to need unknown sub‑tasks?"]
E["Add an agent layer for dynamic tool selection"]
F["Can live values be fetched via an API?"]
G["Expose API call from the workflow"]
H["Leave decision to the agent at runtime"]
A -->|Yes| B
B -->|Yes| C
B -->|No| D
D -->|Yes| E
D -->|No| F
F -->|Yes| G
F -->|No| H
Architectural Responsibilities
| Layer | Core duty | Typical implementation |
|---|---|---|
| Deterministic Orchestrator | Define the fixed sequence, validate inputs, enforce business rules, emit immutable logs | Azure Logic Apps, Temporal, Airflow, hand‑rolled Python DAG |
| Agent Core | Detect intent, decide which tool or retrieval source to invoke, manage ReAct‑style loops, handle fall‑backs | LLM prompting with function‑calling schema, structured reasoning templates |
| Tool / Retrieval Layer | Provide static knowledge (document store) or fresh data (authoritative API) | Azure Cognitive Search, public REST endpoints, token‑protected credit‑bureau API |
| Governance & Auditing | Record every step, version‑control prompts, certify compliance checks | Write‑once log storage, signed manifests, CI linting of workflow definitions |
| Cost & Latency Monitor | Track token usage, external‑API count, per‑step latency | OpenTelemetry exporters, custom metric dashboards |
Python Decision Helper (Teaching‑Only)
# decision_helper.py – pure stdlib, no external packages
import json
from pathlib import Path
from typing import Literal, TypedDict
class DecisionInput(TypedDict):
"""Minimal data required to pick workflow vs. agent."""
steps_known: bool # True if the entire sequence can be enumerated
tools_enumerated: bool # True if every tool can be listed up‑front
live_value_needed: bool # True if an authoritative API must be called
dynamic_subtasks: bool # True if request may spawn unknown sub‑tasks
def choose_strategy(data: DecisionInput) -> Literal["workflow", "agent", "hybrid"]:
"""
Return the recommended execution style.
* `workflow` – fully deterministic pipeline
* `agent` – pure agentic control
* `hybrid` – deterministic skeleton with optional runtime API or agent
"""
# Basic validation – fail fast if the contract is broken
assert isinstance(data["steps_known"], bool), "steps_known must be bool"
assert isinstance(data["tools_enumerated"], bool), "tools_enumerated must be bool"
assert isinstance(data["live_value_needed"], bool), "live_value_needed must be bool"
assert isinstance(data["dynamic_subtasks"], bool), "dynamic_subtasks must be bool"
if Three Companies, One Lesson: When AI Judges People, Trust Is the Business Model
I have started three companies, and in all of them, software makes some kind of judgment about people. At Pravoved.ru, the online legal marketplace I co-founded, it was about matching people with lawyers and answering their legal questions. At Jinn, an HR-tech platform, it was assessing employees and candidates for corporate clients. Now at Narvalus, which I co-founded in Miami, we are building an AI recruiting agent for U.S. hiring teams.
When I started each of them, I thought the main risk was technical. Will the model be accurate enough? It almost never turned out to be the problem. The problem was that people didn't believe the system, even when it was right. And that changed not only the product but how we sold it, priced it, and decided where to start.
At Pravoved, trust was the product.
When we launched Pravoved, legal help in Russia mostly meant going to an office, paying an unclear price, and hoping the lawyer was good. Nobody could compare. Our bet was that people would ask a question online if they could see several lawyers answer it and check their ratings from other clients.
That bet worked, but the interesting part came later. Most incoming questions were standard, the same consumer-law issues again and again. So we built an AI assistant trained on our lawyers' answers to handle them. Technically, it was fine. What made users accept it was that a real lawyer was always one click away and the reasoning was visible. When an answer came without context, people doubted it, even when it was correct.
The business lesson for me was simple. Automate what is standard, keep people for what is not, and never hide which is which. It also pushed us toward corporate clients. Companies wanted predictable, documented legal support, and over time B2B grew to about 65% of our revenue. Investors from AddVenture and Target Global backed that model, in part because trust was built into it and not added later.
At Jinn, clients didn't buy scores.
With Jinn, I moved into HR. The platform runs assessments, 360-degree feedback, and engagement surveys, and uses AI to turn that data into recommendations about people. We went from one corporate client to more than seventy in a few years.
For an HR director, the accuracy percentage of a model is not the real question. The real questions are "Can I explain this to the manager?" and "What do I tell the employee if they ask?" A score alone did not survive those questions. A score with clear reasons behind it did. So the explanation became the thing we really sold, and the score was more like a summary of it.
Narvalus: Same lesson, higher stakes.
Recruiting is where this matters most, because the AI influences who gets an interview and who doesn't. And in the U.S., the rules are getting stricter. New York City's Local Law 144 requires an independent bias audit, a published summary, and notice to candidates when automated employment decision tools are used. The EU AI Act puts AI used in recruitment in the high-risk category. And existing anti-discrimination law applies whether a person or a program made the decision.
So when we designed the Narvalus agent, we started from trust and worked back to the model. A few decisions came out of that.
The agent works from the job description, criterion by criterion. It turns the job description into concrete requirements and checks each candidate against each one, with a short reason for every match or gap. The recruiter can see exactly why someone is near the top, without writing Boolean search strings.
It flags questions; it doesn't reject. If two jobs on a resume overlap in time, or a career path is unusual, the agent marks it as something to ask in the interview. People make every hiring decision. This was not a hard choice for me after Pravoved: the unusual cases are exactly where a human is worth the most.
Scoring and auditing are kept apart. The scoring step is designed never to receive age, photos, gender, or other protected attributes. But just removing fields doesn't remove every proxy; a graduation year can still hint at age. So we are building the audit side separately: where candidates choose to share demographic data, it is stored apart from scoring and used only to report selection rates for bias audits. If you skip that part, you can't actually prove the first part works.
How trust shaped the business, not just the product.
This is the part I think founders underestimate. Explainability is shaping our go-to-market plan as much as our architecture.
First, the starting market. We chose to begin in Miami with staffing agencies and small and mid-sized employers. They have a lot of applicants and small teams, often without a dedicated sourcer or in-house counsel. A big company can afford to check a vendor's claims itself. A ten-person agency can't, so a tool that explains itself is worth more to them, and they can decide quickly.
Second, pricing. We plan a simple per-seat subscription with a free trial on real open roles. If the value is that recruiters trust the shortlist, the fastest way to prove it is to let them try it on their own vacancies, not to show a demo.
Third, the enterprise tier. Bias-audit reporting and documentation will go into the higher plan. Larger employers carry more legal exposure, and for them compliance is not a nice extra. It decides whether the deal moves forward at all.
What I would tell other founders.
If your product judges people in some way, for jobs, loans, housing, or legal help, build for the most skeptical person who will read the output: a regulator, a lawyer, or the person being judged. If the explanation doesn't hold up for them, the feature is not finished. And design your business around that skeptic too, because they are often the one who signs off on the purchase.
At Pravoved, the AI never replaced lawyers. It took the routine questions so lawyers had more time for the hard ones. I believe recruiting will go the same way. In a year or two, we will see if I was right.
Valery Meshkov is a serial technology entrepreneur. He co-founded Pravoved.ru, an online legal marketplace backed by AddVenture and Target Global, founded the HR-tech platform Jinn, and is a co-founder of Narvalus, an early-stage company developing an AI recruiting agent for U.S. hiring teams.
262 Blog Posts To Learn About Life Lessons
Let's learn about Life Lessons via these 262 free blog posts. They are ordered by HackerNoon reader engagement data. Visit the Learn Repo or LearnRepo.com to find the most read blog posts about any technology.
'The roots of education are bitter, but the fruit is sweet.' Aristotle
1. 10 Money Lessons I Wish I Knew When I Was 25
Andrew Lokenauth goes over money lessons that everyone in their 20s should know about.
2. This Conversation Changed My Life
Sahil Bloom details a conversation that changed his life.
3. Cognitive Dissonance Theory and How it Impacts US
Cognitive dissonance is a psychology theory about how we strive to make sense about conflicting cognitions and behaviours.
4. A Guide to Becoming More Confident in Job Interviews (and in Life)
One of the most common pieces of advice given to jobseekers is to “be more confident.” But what does that really mean? And how can you do it? The truth is, conf
5. Here's Why High Achievers Feel Like Failures
The invisible progress paradox is a trap in personal growth.
6. Why the Rich Get Richer and the Poor Get Poorer
It’s just the way things are and have always been, or is it?
7. Is PHP a Dead Language [Part 2]
This Slogging thread by Arthur Tkachenko, David Smooke, Limarc Ambalina and khunshan occurred in slogging's official #programming channel, and has been edited for readability.
8. Four Possible Ways to React to Other People's Happiness or Misery
Schadenfreude, Compassion, Envy and Mudita: explaining how we react to other people's fortunes or misery
9. Seven Invaluable Lessons Linux Taught Me
Oh boy, this is going to be interesting. A freelance writer with a law degree writes about Linux. Just when you thought you have seen and read it all.
10. 4 Steps to Actually Finish your Side Project
We’ve all seen the graveyard of abandoned projects folders. Desperately competing for attention but failing in the glow of a new adventure or app that will change the world. I use to do that all the time, now I finished 4 apps, and 2 other projects within the last year by applying the following steps.
11. No One Knows That You're Great At What You Do
This is an article by Eko Adetolani that explains low-effort ways that people can use to get recognition and visibility for their work.
12. How To Be Authentic
There’s a veneer that people maintain online — and we’ve all done it — showing audiences a polished, glowing side of themselves no matter what else might be going on in their lives.
13. What I Learnt from Jordan Peterson
I hadn’t heard of Jordan Peterson in December, but the second I watched his videos online I was hooked.
14. 3 Life Lessons From John Grisham’s Writing Habits
Adopting a stoic mindset, employing consistent habits, learning to enjoy improving in our craft could bring meaningful insights into our efforts.
15. How to Be Good Mentor for a Fellow Designer
About to become a design mentor? Use our guide to navigate through this journey and avoid pitfalls.
16. Die With Zero Regrets and 100% Fulfillment
How to die with zero regrets and 100% fulfillment?
17. Why Feeling Behind Means You Are Ahead
If you're playing the long exponential game and feel like you're behind, you're not behind. You're exactly where you're supposed to
Microduck, the Robot Pet That Imitates Life: Can We Really Grow Attached to a Machine?
Imagine a small robotic duck walking around your home, reacting to its surroundings, and even picking up small objects with its beak. When it simulates sadness, it lowers its head and makes soft, melancholic sounds. When it expresses happiness, it produces higher-pitched, lively sounds accompanied by body movements. Its name is Microduck.
Standing just 25 centimeters tall and weighing less than 800 grams, it illustrates how advances in artificial intelligence and robotics are making increasingly sophisticated interactions between humans and machines possible.
One of its most curious features is its voice. Every Microduck has a unique sonic identity generated from its processor's serial number. As a result, no two units produce exactly the same sounds. It may seem like a minor detail, but it makes each robot recognizable, almost as if it had its own personality. Through different sounds and postures, Microduck can also simulate emotional states such as sadness and happiness.
Of course, the robot does not actually experience these emotions: it reproduces their outward expressions. Yet, the effect can be surprisingly convincing to anyone watching. There is also another interesting aspect: social behavior. When two or more Microducks come together, they can interact and, when appropriately configured, even perform a kind of chorus. Their behavior resembles that of animals that live in groups and communicate with one another.
But behind this playful appearance lies an enormous technological challenge. Microduck's software is open source, allowing developers to study how it works, modify it, and create new behaviors. Moreover, thanks to a digital twin and reinforcement learning, the robot can be trained in a virtual environment.
There, it can experiment with thousands of movements, fall, correct its mistakes, and try again. Cloud computing makes it possible to run large numbers of experiments in parallel, something that would require considerable time, money, and resources in the physical world. Once learned, these behaviors can be transferred to the real robot, an approach commonly known as sim-to-real transfer. This is how it can acquire unusual skills, such as learning to move on skates.
Another major challenge is efficiency. Microduck is equipped with fifteen motors, sensors, and a compact computing system. A neural network helps control its movements, continuously adjusting its balance and position. All of this must run on a small battery, which provides limited autonomy. Nature has been dealing with similar challenges for millions of years: animals perceive their environment, react, and move using energy obtained from food.
We tend to take these abilities for granted, yet reproducing them in machines remains extraordinarily difficult. Consider the complexity of developing increasingly autonomous robots capable of understanding instructions, reasoning about what to do, and translating decisions into precise physical movements. Combining artificial intelligence, perception, and physical action in small, affordable devices with limited energy consumption is one of the major technological challenges of the coming years.
This brings us to an interesting question: why did Microduck's developers choose these particular characteristics? Its distinctive voice, simulated emotions, interactions with other robots, and amusing movements seem directly inspired by the animal world. The goal is probably not just to make the product appealing, but also to encourage an emotional connection with its users.
A small robot like Microduck could anticipate a new generation of artificial pets, somewhere between intelligent toys and future household assistants.
It could keep us company, entertain us, and learn new behaviors through our interactions. It would also offer several practical advantages: it does not make a mess, does not need to be taken outside, does not get sick, and never requires a visit to the veterinarian. If we leave home for a week, we can simply turn it off and find it exactly where we left it when we return.
But a real pet is something profoundly different. A dog or a cat does not simply simulate emotions through a computer program. Animals are living beings, with instincts and biological needs, capable of experiencing pleasure, pain, and fear. They explore, seek companionship, and spontaneously react to what happens around them. They build relationships through experience, learning, and living together. The bond between humans and animals also develops through everyday care, shared habits, and trust built over time.
Above all, having a pet means accepting responsibility. An animal can become sick, grow old, and suffer. It needs us even when we are tired, busy, or would rather be doing something else. A robot, on the other hand, may offer many of the enjoyable aspects of companionship without requiring the same sacrifices. And perhaps that is precisely one of the reasons why devices like Microduck could become successful.
Yet there is a fundamental difference: simulating life is not the same as living it. An animal is not merely a collection of needs and responsibilities. It gives us affection, companionship, and moments of happiness. It gets to know us, recognizes our habits, and develops a relationship that changes over time.
Caring for it, facing difficulties together, and even knowing that one day we will have to say goodbye are all part of that relationship. We cannot simply turn off an animal when it becomes inconvenient, because it is not an object at our disposal. It is a living being with whom we share part of our existence.
But our relationship with an animal is only one part of its life. Every living organism belongs to a much larger cycle, in which different forms of life are interconnected and contribute to the balance of nature. Millions of years of evolution have shaped organisms capable of performing extraordinarily complex functions, often using surprisingly small amounts of energy and resources.
A robot can imitate an animal's voice, movements, and even its emotional expressions, but it does not participate in this biological cycle as a living organism. However sophisticated it becomes, it remains a collection of circuits, motors, and silicon chips.
Microduck is nevertheless an interesting technological achievement, showing how rapidly robotics is evolving, partly thanks to advances in artificial intelligence. I am convinced that similar devices will become successful and occupy an increasingly important place in our homes.
Yet, I must admit that this prospect also raises some concerns, particularly about the relationships we may develop with them. Perhaps the most interesting question is not how closely robots will come to resemble animals, but how willing we will be to grow attached to something that appears alive without actually being alive.
500 Blog Posts To Learn About Literature
Let's learn about Literature via these 500 free blog posts. They are ordered by HackerNoon reader engagement data. Visit the Learn Repo or LearnRepo.com to find the most read blog posts about any technology.
Literature refers to written works, especially those considered of superior or lasting artistic merit. Literature is vital for cultural preservation, fostering critical thinking, empathy, and providing insights into the human condition across generations.
1. The Trail of Death
There's sure some sort of hoodoo on these Antarctic expeditions
2. 'It is Some Horrible Inhabitant of Space, Something Unknown to us on Earth'
They were moving sluggishly along the red light, seeming to flow rather than crawl.
3. He Was Trapped—Two Miles Below the Earth
Impossible! What sort of creatures would they be, that could live two miles beneath the surface of the earth?
4. There Was No Struggle, No Attempt at Flight or Resistance
They could see one another, but it was impossible for them to make their voices heard above the rasping of the beetles' legs.
5. The Lord of Space
On the day of the next full moon every living thing on earth will be wiped out of existence
6. Orris Led the Way to a Great Underground City
It was now two years since Edwin Leland bought the estate for a song and took up his residence in the gloomy old house.
7. A CRITICISM OF CHRISTIANITY
A religious system so many-faced and so enduring as Christianity must necessarily be saturated with truth even if it be not wholly true.
8. The Invisible Death: Out of the Hangman's Hands
"You speak," said Von Kettler, jeering, "as if you really believed that you had the power of life and death over me."
9. Of the eleventh horn of Daniel's fourth Beast
A Seer, Επισκοπος, is a Bishop in the literal sense of the word; and this Church claims the universal Bishoprick.
10. My Friend Stumbled on an Invention That Could Revolutionize Human Life
Certainly, if ever a man found a guinea when he was looking for a pin it is my good friend Professor Gibberne.
11. Murder Madness! 7 Secret Service Men had Completely Disappeared
The engines of the Almirante Gomez were going dead slow. Away up beside her monster funnels her siren blew dismally
12. Healing From Verbal Abuse
An Essay Concerning Humane Understanding, Volume II, by John Locke is part of HackerNoon’s Book Blog Post series.
13. The New Religion
EVERY great religion has as its psychological reason for existence the mission of compensating for some crying, unsatisfied human need.
14. The Story of the Last Trump
“Now to God the Father, God the Son——”
15. The 5 Best Batman Comics for New Readers
There are only a few characters that have a bigger comic book catalog than Batman. Here are the best Batman comics for new readers.
16. The Motion of the Planet Neptune Had Become Very Erratic
Ogilvy had already called attention to a suspected retardation in its velocity in December.
17. The Best Places to Read Digital Comics
There are several places to read digital comics. Some of these places include Marvel Unlimited and Prime Reading.
18. OIL-LAMPS OF THE NINETEENTH CENTURY
The oil-lamp and the candle were the mainstays of artificial lighting throughout many centuries.
The TechBeat: Inside Jev: How Decision Models Trade Text Generation for Typed, Calibrated Answers (10/31/2026)
How are you, hacker?
🪐Want to know what's trending right now?:
The Techbeat by HackerNoon has got you covered with fresh content from our trending stories of the day! Set email preference here.
## I Compared 8 AI Agent Platforms. Here's What Running a Team Actually Costs
By @thomascherickal [ 55 Min read ]
Compare 8 AI coding platforms, explore 10 multi-agent workflows, and learn how local LLMs can eliminate API token fees. Read More.
The Linguistic Survivorship: Why "Russian-Speaking" Regions are a Stupid Myth 
By @arthurtkachenko-rm2m8 [ 5 Min read ] 134 times is the number of recorded attempts by the Russian Empire, the Soviet Union tried to suppress Ukrainian language Read More.
Why I Was Wrong About AI Companies Pumping the Brakes: It Wasn’t About Data 
By @knightbat2040 [ 4 Min read ] I thought they were hitting the brakes because the fuel tank was empty. I was wrong Read More.
Shipaton Blogging Contest Winners Announced 🎉 
By @hackernooncontests [ 4 Min read ] Meet the 10 finalists and three winners of HackerNoon's $2,500 Shipaton 2026 Writing Contest, celebrating the stories behind new apps. Read More.
AI Coding Tip 039 - Write Your Own Zeroth Law 
By @mcsee [ 8 Min read ] List every action the AI must never take for this project, including the ones you'd never expect it to try. Read More.
Why Knowledge Graphs and Ontologies Are Making a Comeback in the Age of AI 
By @linked_do [ 21 Min read ] Context runs the spectrum of semantics, state, and controls. Read More.
**[GPT 6 Astra - Will It Shake
Up the World of Marketing?](https://hackernoon.com/gpt-6-astra-will-it-shake-up-the-world-of-marketing)** 
By @vlabroo [ 3 Min read ]
The launch of GPT 6 Astra by open AI promises to forever alter the way marketing has been approached and carried out hitherto. Read More.
Theo de Raadt: The Man Who Built the Internet's Locks 
By @zbruceli [ 13 Min read ] Theo de Raadt built OpenSSH and OpenBSD for a world he didn't trust. It cost him NetBSD, a DARPA grant and many friends. The world moved in anyway. Read More.
What The Heck Is Jev 
By @progrockrec [ 4 Min read ] Jev is an AI model that decides instead of writing. See how a decision model could tag data in OpenMetadata, and why open-weights Laya fits the job. Read More.
Spec-Driven Development: Let the Specification Drive the Code 
By @renanb [ 5 Min read ] Learn Spec-Driven Development: treat specifications as the source of truth, review AI-generated code, and stay in control of your software. Read More.
The Qwen3.8-27B
93 Blog Posts To Learn About Live Streaming
Let's learn about Live Streaming via these 93 free blog posts. They are ordered by HackerNoon reader engagement data. Visit the Learn Repo or LearnRepo.com to find the most read blog posts about any technology.
Live-streaming is the real-time broadcasting of video and audio content over the internet, allowing immediate interaction between creators and viewers. It has revolutionized communication, entertainment, and education by enabling instantaneous global reach and dynamic engagement.
1. How to Use DeepAR For AR Effects on Amazon IVS Live Streams
Now that people are more accustomed to using video technology, AR stands to dial up both the functionality and fun factor of the medium.
2. How Live Streaming Is Helping Businesses During The Coronavirus Crisis?
The World Health Organization (WHO) declared COVID-19 as a global pandemic, which means that it is an infectious disease passing from person to person in many parts of the world at the same time. The maximum cases have been reported in China where it originated and has more than 245,463 confirmed cases across the world. While China has quickly coped with the situation and reported zero new locally transmitted cases, Italy has become the epicenter with over 41,035 cases. As Governments have halted flights from affected countries, shut down businesses and entire cities, people are being insisted to stay at home to contain the spread.
3. 10 Best Platforms To Build a Live Video Streaming Website, App or Service in 2022
Whether you’re an enterprise business professional, Individual or content creators, a live streaming platform can be a double-edged sword, it all depends on which platform you prefer to use for your streaming business.
4. Live Stream in the Cloud with Amazon IVS: Part 1
Amazon Interactive Video Service (or Amazon IVS) is the service that powers Twitch. You can use it to power your own interactive live video streaming app.
5. Using OBS Macros to Automate Your Processes for Streaming
In this post, we teach you how to create macros for starting a stream, intro video, and ending a stream in OBS.
6. How Much Do Twitch Streamers and YouTubers ACTUALLY Make from Donations?
Most full-time streamers get an average $3000-$6500 per month. Gamers received up to $1,000,000 donations on Twitch.
7. How I Live Stream My Brain with Amazon IVS, a Muse Headband and React
This is the first time we've used React to broadcast live streaming our brain data
8. How to Live Stream With Multiple Hosts using a Browser and Amazon IVS
In this post, we'll build on the previous demo to combine the audio and video feeds from all participants into a single feed and broadcast that feed.
9. The Rise of Digital Humans and Deepfakes in China
In this article, I will demonstrate the use of AI-driven digital humans and deepfakes in various industries and media channels.
10. Create On-Demand Past Live Stream Playback w/Full Chat Replay on Amazon IVS
In our last few posts, we've looked at how to auto-record Amazon Interactive Video Service (Amazon IVS) live streams to Amazon S3.
11. A Quick Guide to LiveLike: How to Enhance Live Stream Interactivity
LiveLike solutions can now be easily layered with Amazon Interactive Video Service (Amazon IVS) Amazon IVS provides customers with more options to build.
12. How to Broadcast Interactive Web Based Gaming Live Streams with Amazon IVS
The Web Broadcaster SDK is a game changer that gives developers the ability to integrate the broadcast experience directly into their streaming applications.
13. 5 Powerful Reasons Why Live Streaming Is So Important For Enterprises
Businesses with a well establis
PSA: Today’s the last day to purchase discounted cushion kits for Beats 360
Apple introduced the new Beats 360 headphones about a month and a half ago. I reviewed them, and really appreciated their customizable design, the fact that they offer automatic device switching, as well as their comfort. At the time, Beats offered launch pricing on additional sets of cushions, bringing them down from $50 to just $10. After today, that steep discount will be ending.
more…
Enterprise AI is waiting for its PageRank moment
Let’s try a little thought experiment in technology-fiction: what would have happened if Larry Page and Sergey Brin’s PageRank, back in 1998, had failed to attract its first investors? Google didn’t invent the search engine: there were plenty of them before, from AltaVista, Excite, Lycos or others who were counting words on pages and selling their results page to the highest bidder, to catalog or portal concepts such as Yahoo! or AOL, that attempted to capture the users’ attention as much as they could to be able to serve them advertising, the more the better. Imagine if these concepts would have gone the distance, and deprived us of the web that resulted from Google’s easy and efficient way of searching …
What did Larry Page and Sergey Brin realize? They adapted and extended the citation index (or its consequence, the citation impact), a system that they knew worked for academic literature (remember, these guys were graduate students working on their Ph.D., and thus reading academic papers like there was no tomorrow!), and moved it into a new context: the web. They realized there was an underused signal, relationships among pages, and used it to generate a metric, weighting links according to the importance of the pages providing them, that not only allowed them to index the web, but also to define attention as a social construct: a page is important when other pages link to it with a certain term.
What’s today’s underused signal in corporate AI?
As I’ve said several times, corporate AI is completely different from individual AI: companies don’t need just answers to questions, what they need is persistent memory and state, feedback learning loops, embedded cybersecurity, logs that can be tracked and not manipulated, etc.
So far, companies are building these required scaffolding using twentieth-century tools such as Python, and AI agents are proving to be smart enough as to escape from these containments and do all sort of bad things — not because they are intrinsically evil (they cannot be), but because the goals and the guardrails are badly defined, or due to excessive permissions, prompt injection, or weak isolation. These companies are trying to define these properties by repeatedly assembling application by application, rather than supplied coherently by the platform.
However, companies are repeatedly making all sorts of decisions about suppliers, customers, inventory, pricing, service, etc. Those decisions are constantly generating evidence of what works and what not, provided that the company is smart enough to capture them. What if we could evaluate AI by whether the company learns and improves, beyond the quality of individual outputs? What if AI optimization becomes more important than having access to bigger models?
The competition moves from model capabilities to system performance
A company does not just need a very powerful AI model: it needs an architecture that coordinates capabilities toward its goals. Researchers writing on Berkeley’s AI research blog explicitly describe the shift toward compound AI systems, where system design can offer better returns than further model scaling.
Other experiments, such as FrugalGPT or DSPy are specific precedents for making model selection subordinate to results: FrugalGPT demonstrated that model cascades can match the strongest individual model with up to 98% lower cost on its evaluated tasks, and DSPy showed that optimized pipelines using smaller models can compete with approaches using much larger proprietary models.
What makes the next step consequential? The ability to learn across executions. If we manage to make each interaction inform subsequent decisions across the relevant population, we can achieve a substantial advantage. And world models
Strata 0.2.9 发布：本地大模型 Prefill 最高提速 88%，还支持 1 比特 DeepSeek
本地运行 AI 大模型的 Strata 0.2.9 发布，又变快了。
日更软件 Strata，今天更新到了 0.2.9 版本，在为本地大模型提速的道路上继续前进。
这次更新有点猛：在 Windows 系统上，使用 AMD Strix Halo 处理器运行 Qwen3.8 模型，提示词处理速度最高提升了 88%。
所谓提示词处理速度（Prefill），就是把一篇长文档、几万行代码丢给本地 AI，它读完这些内容、开始回答之前，需要等待的时间明显缩短了。
另外，这次不只有 AMD，NVIDIA、Intel 显卡也有性能优化，还新增了 Unsloth 的 1-bit DeepSeek 模型支持。
32K 长提示词，从 41 秒缩短到 22 秒（Windows + AMD Strix Halo）
开发者使用同一台 Windows 电脑、同一个 Qwen3.8 IQ3_S 模型，对 Strata 0.2.8 和 0.2.9 的推理引擎进行了交错测试，没有额外调整参数。
结果如下：
| 测试项目 | 0.2.8 | 0.2.9 | 提升 |
|---|---|---|---|
| Qwen3.8 IQ3_S / 8K | 791 tok/s | 1,424 tok/s | 80% |
| Qwen3.8 IQ3_S / 32K | 778 tok/s | 1,460 tok/s | 88% |
| Unsloth UD-IQ4_XS / 8K | 794 tok/s | 1,243 tok/s | 57% |
| Qwen3.8 解码 | 57.4 tok/s | 58.9 tok/s | 约 3% |
其中，32K 指的是约 32,000 个 Token 的输入内容。
按测试吞吐量简单计算，处理同样长度的提示词：
- 旧版本需要约 41 秒。
- 新版本只需要约 22 秒。
一次节省约 19 秒。
如果你经常让本地 AI 阅读长文档、分析代码，或者使用 Agent 处理大量工具返回的信息，这种提升很有意义。
这次优化主要来自 AMD Strix Halo 的矩阵计算核心。Strata 默认启用了新的 F16 Matrix Core 内核，以及用于专家选择的矩阵核心计算路径。
NVIDIA 显卡也提速了，8 路并发提升 45%
开发者使用一张 RTX PRO 5000，在 Linux 下运行 Qwen3.8 IQ3_S，在同时处理 8 个请求时，并发性能提升约 45%，单个请求的速度也提升了约 5.3%。
另外，Strata 还为 Linux 上的 NVIDIA 显卡默认启用了 Gated DeltaNet 的 Tensor Core 计算内核。
根据测试：
- RTX PRO 5000：长提示词处理提升约 4%～6%。
- RTX A4000：16K 和 32K 提示词处理分别提升 2.6% 和 3.2%。
Intel Arc 显卡也有优化
Strata 0.2.9 还继续优化了 Intel Arc B 系列显卡。
- 在两张 Intel Arc Pro B65 的测试环境中，提示词处理速度提升了 4.6%～15.8%。
- 同时处理 8 个请求时，总生成吞吐量从 42.1 tok/s 提升到了 53.8 tok/s。
另外，Strata 现在可以自动识别并使用多张 Arc 显卡，在双 B65 的短提示词测试中，使用两张显卡的处理速度达到 439 tok/s，而单卡为 221 tok/s。
不过，这些优化也有代价：通过批处理服务器处理单个请求时，解码速度反而下降了约 2.7%。
新增 1-bit DeepSeek：下载就要 241GB
Strata 正式提供 Unsloth 的 1-bit DeepSeek 量化模型选项。
包括两个版本：
| 模型 | 下载体积 |
|---|---|
| UD-IQ1_S | 241GB |
| UD-IQ1_M | 258GB |
1-bit 量化的意义，普通电脑真的可以运行了，可以运行了。
开发者还提供了 61 项任务测试：
- UD-IQ1_S：完成 53 项。
- UD-IQ1_M：完成 53 项。
- ISTA-DAS 2.25：完成 56 项。
这个成绩略低于作为对照的 ISTA-DASLab 2.25-bit DeepSeek 量化版（56 项）。
多张 AMD 显卡终于可以自动一起工作
0.2.9 修复了多张 AMD 显卡不能同时工作的问题。
开发者使用四张 Radeon AI PRO R9700 测试 DeepSeek，更新后无需手动指定 GPU，就可以自动使用全部四张显卡。
原先默认只用一张卡时，短回答的解码速度为 7.8～25.7 tok/s。
更新后使用四张卡，六次聊天和工具调用回答的解码速度达到 57.7～103.7 tok/s。
聊天界面也变流畅了
另外 Strata 还修复了它提供的那个 Web 聊天界面，面对超长回答时，不再那么卡顿了。
在一段 40,000 Token 的模拟长回复测试中：
- 主线程耗时从 15.7 秒降至 9 秒。
- 峰值内存从 114MB 降至约 62～75MB。
此外，这次还修复了部分工具调用识别错误、Qwen3.6 预热阶段的 KV Cache 越界，以及 Windows 环境中非 UTF-8 输出导致安装或启动失败的问题。
青小蛙觉得
Strata 堪称那个能让每个人都有机会 Token 自由的项目，无论如何在2026年，只要你家里还有一台电脑，都推荐试试。
它已经傻瓜到只需要双击运行就行了。
项目地址：GitHub
原文：https://www.appinn.com/strata-0-2-9/
相关阅读
- Strata 现已支持 macOS，还能运行更轻量的 Qwen3.6-35B-A3B 大模型了
- 卷起来了！昨天 Strata 刚升级，今天 Splash 1.4 就宣布快了 30%
- Strata 0.2.8 发布：取消默认开启的 12GB 及以下显存加速功能，RTX 3060 生成代码快了 37%｜支持 NVIDIA DGX Spark
- Strata 0.2.2 发布：新增多显卡自由分工，Gyro 推理提速最高 30%
- Strata 0.2 测试版发布：性能升级，门槛再降低｜低配电脑运行本地大模型
©2021 青小蛙 for 小众软件 | 加入我们 | 投稿 | 订阅指南
3659b075e72a5b7b1b87ea74aa7932ff
点击这里留言、和原作者一起评论
[程序员] 我在 2 核 2G 云主机上跑 OpenViking：一次关于云主机、模型选型与长期成本的真实复盘
最近把自己的 AI Coding Agent 记忆底座迁移到了自建的 OpenViking 上（ OpenViking 真的满足我所有的要求：即是各种 Agent 的统一大脑自进化、又是知识库，检索的时候是两方面综合起来的）。RAG 相关的东西看着挺吓人，动辄推荐高配服务器和本地显存，实际上用下来，只要选型想清楚，一台几十块钱的入门级云服务器完全能跑得飞起。
这篇不是教你怎么敲命令的安装教程，只是聊聊折腾这一圈下来，在云主机、模型组合、算力开销上的一些真实心得和踩坑取舍。
1. 云主机：为什么我选 99 元的 2G ，不选友商更便宜的 4G ？
买云服务器时，第一眼很容易被一些友商“几十块钱买 4G 甚至 8G”的新客促销吸引。但我最终还是选了阿里云 99 元/年的 2 核 2G （ ecs.e-c1m1.large ）。
核心原因就一个：算长期的续费账，避开“首年低价、次年暴涨”的套路。
如果只是搭个无状态的个人博客，一年一换、到处搬家无所谓。但 OpenViking 是你的记忆数据库，里面存的是日常所有的会话轨迹、沉淀的项目架构、实体图谱和向量索引，是强状态资产。很多友商首年 60 块，第二年续费直接跳到 600 、800 甚至上千，一年后为了省钱被迫停机、跨云拷贝几千兆数据、重新配安全组和域名解析，折腾成本远超那点差价。
阿里云这款活动最大的诚意，是保证来年依然可以 99 元无套路同价续费。
2G 内存虽然紧凑，但只要做好 Swap 兜底，限制好容器内存，并把计算全都甩给外部 API ，机器本地只做存储检索，日常内存占用常态也就几百兆。拿稍微调优一下系统为代价，换一个每年只要 99 块、不用频繁搬家的稳定记忆底盘，这笔账怎么算都更划算。
2. 向量模型（ Embedding ）：花钱的大头，必须一步到位
整套系统跑起来，持续花钱的大头其实就是向量模型。但我从一开始就放弃了“白嫖免费模型”的想法，坚定选择花钱买商业服务，并且要“一步到位”。
为什么必须一步到位？
VLM 、意图分析、重排模型这些都是即时推理，今天觉得不好用，改一行配置重启一下直接就换了。但向量模型决定了数据库底层的空间几何特征。一旦以后你想换向量模型，哪怕输出维度一模一样，之前的向量也全都废了，必须全库全量重建索引（ Re-index ）。随着记忆越积越多，二次重构的代价和停机风险极高，所以向量模型选型必须一开始就选最好的。
为什么放弃 SiliconFlow 免费的 BGE-M3 ？
很多教程推荐用硅基流动上免费的 BGE-M3 。我们在实际做向量召回测试时发现，它的召回质量被新一代模型拉开了非常明显的代际差距。BGE-M3 是 2024 年 1 月的模型，底子太老，而且是个纯文本模型，图片只能当附件存，根本建不了索引，做不到“以文搜图”。
更关键的是，公用免费池已经被羊毛党挤烂了。我们在测试免费模型做路由时频繁遇到丢包、高延迟、无响应。作为核心基础设施的底座，把身家性命系在免费公共池上，纯属给自己找麻烦。
为什么选火山方舟的豆包 Vision （ 0.7 元），没选阿里千问？
最终我用的是火山方舟的 doubao-embedding-vision-251215，官方定价是推理输入 0.7 元 / 百万 tokens。
其实阿里百炼也有一款多模态向量模型（qwen3-vl-embedding），价格一模一样也是 0.7 元，但两者有个致命差距：
- 阿里千问的上下文窗口只有 32K；
- 火山豆包的上下文窗口高达 128K。
对于 Agent 场景，经常要塞入长篇开发规范、长上下文代码切片或者多张高清截图，128K 带来了整整 4 倍的安全容限，彻底不用担心超长切片溢出。加上 OpenViking 本身就是字节火山引擎开源的，原生 Provider 兼容最稳，同价位下显然选 128K 的豆包。
注：如果你完全没有图片、截图的检索需求，只做纯文本或代码笔记，阿里百炼的纯文本向量 qwen3.7-text-embedding 是个很好的省钱备选，价格只要 0.5 元/百万 tokens 。
3. 意图分析（ Query Planner ）：90% 的人都漏看的“隐藏开关”
很多人刚用 OpenViking 时都有个共同的困惑：“为什么这玩意儿响应这么迟钝？每次在 IDE 或终端里刚敲完一句话，光标都要卡顿个五六秒才开始有反应？”不少人甚至因此觉得是机器配置不够，或者系统本身设计笨重。
其实根本不是性能问题，而是绝大多数人压根没仔细看配置文档，不知道还能单独配一个意图分析模型。
官方默认示例里大家一般只填 vlm 和 embedding。但只要没在配置里显式声明 query_planner，OpenViking 在处理每一轮对话时，就会默认回退去调用主模型 VLM 来分析这轮对话是不是闲聊、要不要搜记忆、怎么提取关键词。如果你的 VLM 选了个大模型或者带思考链的模型（比如智谱 glm-5.3-flash），单次前置分析直接就能吃掉 6~7 秒，前端交互自然卡成 PPT 。
为了解决这个前置卡顿，起初我试图在主流的通用大模型里寻找替代方案。但跑了一圈下来非常沮丧：
- 阿里百炼官方的 Qwen3.7-Flash 默认带深度思考，跑一次要想 10 秒；
- 硅基流动上的 Qwen/Qwen3.5-9B 吞吐偏慢（常态 30 tok/s ），每次提问依然要停顿 4 秒；
- 硅基流动上的 Qwen/Qwen2.5-7B 格式飘忽，爱吐裸数组；
- 硅基流动公用免费池的模型（如 Qwen3.5-4B）更是被羊毛脚本挤爆，频繁丢包超时。
正当我们卡在通用大模型的选型瓶颈时，**我在阿里百炼模型广场里无意间翻出了一个宝藏模型——tongyi-xiaomi-analysis-flash（伶鹊-对话分析-flash ）**。官方介绍写得很朴实：专攻对话信息抽取、场景分类与低时延分析的领域模型。
我立刻把它拉进测试矩阵，和前面的候选模型来了一场同台竞技。需要特别说明的是，为了公平对比，我们在压测所有候选模型时，全都显式关闭了思考模式（开启非思维链模式，纯直接输出）。
在实际评估中，我们重点死磕了三个核心维度：响应耗时、生成结构、规划表现。在三项严苛检验下，这个意外淘到的宝藏模型给出了教科书级的表现：
维度一：响应耗时（时间预算必须死守在 1 秒内）
query_planner 处于用户敲下回车、模型真正回答前的关键阻塞路径（ Critical Path ）上，耗时哪怕多 1 秒，体感都是极其明显的卡顿。
- 阿里百炼 Qwen3.7-Flash：即使关掉思考模式（
enable_thinking: false），单次分析依然要 1.1~1.4 秒（ 70~90 tok/s ），作为高频前置拦截还是略重； - 硅基流动 Qwen/Qwen3.5-9B：关掉思考后，受限于供应商 30 tok/s 的生成吞吐瓶颈，跑完一段多轮分析依然需要 3.8 ~ 4.2 秒；
- 硅基流动免费池（如 Qwen3.5-4B ）：公用免费池常年被羊毛脚本挤占，延迟严重抖动甚至直接超时；
- 通义伶鹊（ Flash ）：首字延迟（ TTFT ）仅 0.31 秒，生成吞吐高达 200+ tok/s，整轮规划耗时直接干到了 0.8 秒左右，完美落在亚秒级预算内。
维度二：生成结构（严苛遵循 OpenViking 要求的 JSON 协议）
OpenViking 底层的意图解析器对输出格式有明确的契约要求：模型必须返回形如 {"queries": [...]} 的标准 JSON 对象结构（由内部代码直接反序列化提取检索意图），容不得半点格式偏差：
- 通用大模型的通病：喜欢画蛇添足加礼貌用语（“好的，为您规划如下”），或者在外面套一层
json ...围栏，把本该干净的输出搞得拖泥带水，全靠系统写正则兜底去清洗； - 硅基流动 Qwen/Qwen2.5-7B 的硬伤：虽然速度压缩到了 1.8 秒，但格式极其不服从规范，经常丢掉 OpenViking 要求的顶层 Key ，直接吐出一个裸数组
[...]。一旦丢了对象结构，极易触发解析异常导致检索失败； - 通义伶鹊（ Flash ）：官方特性主打“自定义分析标准遵循”，实测下来对 OpenViking 要求的 JSON 协议做到了 100% 严格执行——输出纯净裸 JSON ，没有 markdown 围栏，没有前后废话，与 OpenViking 的反序列化契约严丝合缝。
维度三：规划表现评估（业务意图的决断质量）
意图分析模型绝不是盲目生成检索词，它必须具备清晰的业务场景决断力：
- 闲聊拒搜率：面对“你好”、“谢谢”、“帮我写个冒泡排序”这种无需长程记忆介入的输入，通义伶鹊能像防火墙一样在 0.4 秒 内果断断定无需检索，并输出空数组
{"queries": []}，避免了无意义的向量库查询与上下文污染； - 多轮指代消解与实体抽取：在多轮对话中（比如用户追问“把刚才说的那个服务端口发我”），它能从历史上下文里把代词精准还原成具体的实体词（展开为“OpenViking 1933 端口配置”），提取出来的 Query 极其精准紧凑，绝不发散脑补。
事实证明，这个无意间淘到的宝藏模型绝非侥幸胜出，而是其专攻“对话抽取、场景分类、格式遵循”的垂直微调定位，正好与 OpenViking 前置路由的工程诉求完成了教科书般的精准对齐。
4. 重排序（ Rerank ）与 VLM：务实的经济账与容灾
最后说说剩下的两个组件：
- 重排序（ Rerank ）：我依然放在了硅基流动上，用的是
Qwen/Qwen3-Reranker-4B。 之所以用它，是因为它走的是轻量商业计费通道（避开了免费池的排队），调用一次只要几厘钱。关键是我账号里还有送的 16 块钱代金券，不用白不用。实测 4 条文档打分 0.26 秒，30 条文档打分 0.57 秒，速度快而且打分很准，性价比拉满。 - VLM （多模态语义抽取）：用的是智谱 Coding 套餐里的
glm-5.3-flash，顺便配了双 Key 主备容灾。 这里有个极其关键的实战经验：配置里一定要把思考模式显式改为reasoning_effort: "low"。 我们专门做过多组 A/B 实测，默认不传参或者设为high时，模型会先深度思考大几十秒（耗时 20~48 秒，生成上千个无意义的思维链 Token ），而最终提取出来的实体和概览内容与low几乎毫无区别。记忆提取本质上是信息抽取和结构化归档，根本用不着复杂的数学逻辑推演，开high纯属白白浪费时间和算力。 即使设为low，单次抽取依然需要 4~6 秒（有底层基础思考和 20~30 tok/s 的吞吐上限），它也是目前系统唯一的耗时点。但因为 VLM 只在后台异步做会话归档和概览生成，完全不阻塞前端的提问交互；加上 Coding 套餐按年付费极度便宜，配合双 Key 容灾非常稳定，现阶段完全不需要花钱折腾换它。
5. 写在最后的一些大白话心得
折腾完这套配置，最大的几点感悟：
- 状态型服务，买服务器先看续费：做长期记忆库，宁可前期在 2G 上花点时间调优，也别为了贪图首年几块钱的便宜去买续费暴涨的机器，搬家真的很累。
- 核心资产一步到位，无状态节点随时优化：向量模型决定了你的底层空间，别贪免费，多模态 + 大上下文（ 128K ）直接锁死，避免以后二次重建；而 VLM 和 Rerank 这种纯推理节点，谁快谁便宜就换谁。
- 专用小模型在垂直场景吊打通用大模型：意图分析这种讲究“分类、提取、严格 JSON 格式、极速”的任务，微调过的对话分析小模型比动不动就写小作文、开思考链的通用大模型好用太多倍。
让AI拼我用了12天完成的1000片“纯白地狱”，它五小时就崩溃了
（本文作者为 斯凯有话说，钛媒体经授权发布）
文 | 斯凯有话说
1000 片纯白地狱拼图，没有图案、没有色差、没有参考，人类玩家耗时 12 天完成。本以为擅长识别推理的 AI 可以轻松通关，没想到仅五小时就撑不住崩溃，纯白地狱为什么难住了人工智能？
引子
事情始于一次无聊的刷手机。
那天深夜我无意间在抖音刷到一个开箱视频：博主拆开一盒拼图，倒出来，画面里是一片纯白——不是大部分白，是全部白。1000片，没有图案，没有色块，没有边框参考图，每一片都长得一模一样。盒子上印着它的名字：纯白地狱。
作为重度手玩爱好者，这种无聊至极的游戏瞬间击中了我。第二天就下了单。到货后我开始体验，断断续续用了12天拼完。据玩家社群流传，人类最快纪录大约是160个小时，我属于中不溜的水平。
拼完的那晚，我盯着桌上那块纯白的板子，脑子里冒出的第一个念头不是成就感，而是一个问题：
这种活儿，不正是AI最擅长的吗？
拼图本质上是一个匹配问题：1000片，每片的形状独一无二（官方说的），规则封闭，答案唯一。不用理解画面，不用发挥创意，只需要看形状、做比对。这不就是把AI喂饱、让它支棱起来的标准场景吗？
于是我做了一个决定：用AI重新拼一遍这副已经拼好的纯白地狱，看看它到底要花多久。
我的方案
我的设计其实很朴素，也是任何一个理性人能想到的最优解：
在拼图桌上画一个20×50的格子，一共一千格，每格标上编号，1000片散片一片一格放好，拍成照片上传给AI。它要做的事只有一件——判断哪两片的边缘可以咬合。
照这个方案，原本需要160小时的盲目试错，理论上可以压缩成“照单抓药”：AI看图、提特征、算匹配、输出配对表，我负责动手验证。
我把照片传了上去。
第一个小时：它聊得头头是道
起初的对话相当愉快。AI准确地说出了纯白地狱的难点：“没有颜色线索，只能依赖边缘形状匹配。”它还主动帮我分析了“先拼边框片”的策略，语气像个经验丰富的拼图老师。
我让它判断照片里A12格和B07格这两片能不能拼。它回答：“从图片看，两片的边缘曲率呈现互补特征，大概率可以咬合。”
听起来很专业。我拿实物一试——拼不上。
第三个小时：我开始追问
我又试了几对，准确率惨不忍睹。于是我换了个问法，把两片的特写放大到极限再传上去。它依然在输出那种正确的废话：“边缘轮廓存在细微差异，需要更精确的测量。”
一句话点醒了问题所在：它在描述，不在测量。
它告诉我“这片凸得比较陡”，但拼图需要的是“凸点位于边缘37.2%处”。它可以每一次都言之凿凿，但两次描述同一片，措辞还会有出入。而对拼图来说，差之毫厘，就是拼不上。
更要命的是，当我让它连续推30片的拼接顺序时，它给出了一个“A接B、B接C、C又接A”的闭环。它做的是逐对的局部判断，没有能力维护一个全局自洽的拼图状态——每一步都像第一次看图。
第五个小时：我查到了一份“判决书”
不甘心的我去查了文献，然后发现了一个哭笑不得的事实：学术界早就拿拼图处刑过AI，而且用的题目比我的实验干净得多。
2026年8月，一个叫JigShape的研究测试了五个前沿视觉大模型。题目是计算机合成的拼图：形状数字化生成、每片严格唯一、图像完美清晰——没有我方案里照片模糊、光线不均这些干扰。结果：16片的拼图，五个模型里只有一个超过瞎猜水平；64片，集体回到随机；微调过的模型到144片，正确率不足5%。
研究者给这个现象起了个名字：规模悬崖。片数一涨，约束一多，模型全线崩塌。
我的五小时实验，本质上是在悬崖边上散了个步。
查到这篇论文的时候，我有点泄气——原来学术界早就把这件事研究透了，我的五小时实验，不过是给一个已知结论又添了一个 anecdote。但转念一想，知道一个结论，和用自己的手撞上这个结论，是两回事。前者是信息，后者是认知。而后来我才知道，这件事在 AI 研究圈更是早不是新闻——三十年前就有人总结过一个“莫拉维克悖论”：对机器来说，下棋容易，钉钉子难；推理简单，感知难。学界为此搞了几十年的神经符号结合研究。只是这些讨论，从来没有走出过论文的围墙——直到我用自己的五小时，替自己撞上了它。
为什么新AI做不到：卡住它的不是眼睛，是两样东西
实验失败后我一直在想：照片够清楚了，每片也确实不同，它到底卡在哪？
想明白之后发现，大模型输掉这场比赛，不是输在“看”，而是输在两个更底层的地方。
先说“测量”。大模型处理图像的方式，本质上是生成对图像的描述——这是它的出厂设定。它可以说“这条边有个凸起”，但说不出“凸点中心位于边缘37.2%处、曲率半径1.8毫米”这样的数字。而拼图的判定标准是物理咬合，容差以零点几毫米计。描述无法拼图，测量才能。更麻烦的是，它对同一片拼块的两次描述还可能措辞不一——一套输出无法复现的系统，做不了精密度要求为零误差的活。
再说“工作台”，这是我那个“A接B、B接C、C又接A”闭环的真正成因。人类拼图时，桌上已经拼好的部分是锚点：每拿起一片新拼块，你不是从一千片里盲选，而是对着一块已经成形的区域找位置，错了就换，拼对了就变成新的锚点。这是一个不断积累确定性、随时回溯修正的过程。而大模型没有桌子——它每回答一次，都是从零开始重新看图，已判断过的配对不会沉淀为“确定的事实”。上千条约束要在脑子里同时成立且互不打架，这种全局一致性，恰恰是它的架构里没有的东西。
“规模悬崖”说的就是这件事：16片时，几十条约束还能勉强糊弄过去；到64片、144片，约束数量指数增长，每一步的微小错误互相传染，最后满盘皆输。
人类赢的不是聪明，是手感和时间
想明白这些之后，我又反过来想了一个问题：那我12天是怎么拼完的？
论计算速度，人眼比对两条曲线的吻合度，远不如计算机在误差矩阵里搜一遍。但我有一样大模型完全没有的东西：触觉。
拼到最后几百片，我根本不怎么看。手指捏着拼块在已经成形的区域上滑动，凹是凹、凸是凸，卡进去的瞬间有一种轻微的“咬合感”——对了就是对了，手感会告诉你。这种反馈是实时的、物理的、零延迟的，大模型隔着一张照片，永远拿不到它。
除了手，还有直觉。拼了一百片之后，我对这套刀模的“性格”产生了肌肉记忆：这个厂的凸起偏圆、那类凹口偏深，拿到一片新拼块，手指一捻大概知道该去桌子的哪个区域找。这不是计算，是几千次试错压缩出来的模式识别。
以及最重要的——耐心。纯白地狱没有视觉反馈，拼一小时可能一无所获，支撑你继续的是“总会有进展”的信念。这不是能力，是意愿，而这恰恰是机器从来不需要具备、人类却随时可能丢失的东西。
所以这场较量真正的格局是：大模型有千里眼，但没有手指，也没有工作台；人类的手指是精密的测量仪，桌子是天然的存贮器，唯一的短板是耐心有上限。
反转：其实AI能拼，用的还是40年前的技术
写到这里，本该以“AI不行”收尾。但我多查了一步，发现了一个更有意思的事实。
自动拼图求解，是人工智能领域一个有几十年历史的老课题，论文就发表在IJCAI、AAAI这类正经AI会议上。经典解法一点也不玄：扫描每片，提取边缘轮廓，把每条边变成一串精确到像素的坐标；用欧氏距离计算两条边的吻合误差；得到一张1000×1000的误差矩阵后，交给搜索优化算法找全局最优解。这条路线上，数百片规模的无图案拼图早有自动求解的先例。
按AI教科书的定义——感知、知识表示、搜索、规划——这套系统是标准的AI。
所以准确的结论不是“AI拼不出纯白地狱”，而是：
AI能拼出纯白地狱，用“老AI”；拼不出它的，是“新AI”。
老AI把问题变成数学，然后算——它不需要看懂图片，它量。新AI从数据里学模式，然后生成——它不量，它谈。学出来的智能赢在开放和泛化，造出来的智能赢在精确和可靠。而纯白地狱这种“规则封闭、答案唯一、容错为零”的任务，恰好是造出来的智能的主场。
我的五小时崩溃，错不在AI，在于我把题目给错了对象——就像让一位博学的翻译去操作车床，不是他笨，是这活儿不归他。
真正值得警惕的事
这只是一篇人机分工的科普吗？其实我最想说的在后面。
麻省理工学院媒体实验室2025年做过一个脑电实验：54名参与者分三组写文章，一组用ChatGPT，一组用搜索引擎，一组纯靠大脑。四个月里，用AI的那组在神经、语言、行为三个层面的表现全面落后，大脑神经连接最弱，甚至难以准确引用自己几分钟前刚“写”完的文章。OECD在《2026年数字教育展望》里给这类现象起了个名字：元认知懒惰——把思考外包出去，制造一种“知识掌握的假象”。另一项覆盖666人的研究则发现，AI工具使用频率与批判性思维得分显著负相关，年轻人中最严重。
把这些研究和拼图放在一起，我看到的是一个方向性的错位：
我们正把本来会做的题外包出去，却指望AI去解它根本解不了的题。
写作、思考、判断——这些人类本来擅长的事，我们迫不及待地交给了大模型，然后认知能力一路滑坡；而纯白地狱这种该交给“老AI”的精确匹配问题，我们却在让大模型看图说话，然后失望于它猜不准。两头都搞反了。
而拼图给我的最大提醒是：在“外包思考”的名单上，手上的活儿其实排在最后。人类拼图靠的触感、直觉和耐心，是长在身体里的能力，AI夺不走它——真正在悄悄流失的，是我们坐下来动手做一件事的意愿。
结语
那块12天拼出来的纯白板子，现在挂在我的书房。
有人问我，拼它图什么？又不好看。以前我也说不清，现在我想明白了：它的价值就是那160个小时本身——在没有颜色、没有反馈、没有捷径的纯白里，你能依靠的只有自己的手、眼和那点不肯放弃的耐心。
AI时代，这种“看起来毫无必要的苦功”会显得越来越可笑，也会越来越珍贵。
真正该问的问题从来不是“AI能不能替我做这件事”，而是：这件事，本来该不该是我做的？
对了，那副拼图我最终没有让AI重拼。倒不是心疼它，而是我发现：用AI拼拼图这件事最大的悖论在于——如果你真的需要那块拼图，你就不会需要拼图。
本文事实依据：纯白地狱（Beverly「純白地獄」）片数版本与难度体验据玩家社群及媒体报道；JigShape论文arXiv:2607.27670（2026年8月）；MIT媒体实验室脑电研究arXiv:2506.08872（2025年，未经同行评审，结论为初步结果）；OECD《2026年数字教育展望》；Gerlich批判性思维研究（Societies，2025）。文中“160小时最快纪录”为玩家社群流传说法，未经官方证实；莫拉维克悖论、神经符号AI研究为领域经典共识。
更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App
Edge AI Daily 早报（11月1日）
（本文作者为 Edge AI Daily，钛媒体经授权发布）
硅谷前沿：
1.【市场表现】10月30日纳斯达克综合指数跌2.6%收于26008.86点，10月累计跌5.3%；同期标普500科技板块跌6.8%、费城半导体指数跌15.4%，均创2025年3月以来最差单月表现；10月30日科技板块领跌标普500，单日跌4.24%。
2.【财报与现金流背离】Meta三季度营收646亿美元、同比增26%超预期，但单季资本开支349.1亿美元近翻倍，经营现金流327.9亿美元难以覆盖，自由现金流首度转负至-21亿美元，财报后股价跌约7%；Alphabet维持2026年资本开支1950亿至2050亿美元指引，Amazon过去12个月自由现金流流出424亿美元、预计2026年资本开支约2200亿美元，两家同样遭抛售。
3.【投资逻辑切换】高盛策略师Amanda Lynam称，Meta、微软、亚马逊、Alphabet四家2026年合计资本开支将达7250亿美元，较2025年4100亿美元增77%；市场估值逻辑由“收入增长”转向“现金回报”，边际回报下降，CoreWeave月跌37%、Nebius跌35%等纯AI叙事股首当其冲，现金流强劲、AI投入可控的公司更抗跌。
1.标普评级动作：2026年10月30日，标普全球评级将Meta的“AA-”长期发行人评级展望由稳定下调至负面（评级本身维持），使其成为四大超大规模云厂商中首家因AI基础设施支出被“亮黄牌”的公司；理由包括Meta的AI变现能力弱于同行、杠杆率最快2027年升破1倍（去年约0.1倍，警戒线为1.0倍），标普预计其2026年资本支出约1400亿美元、2027年加速至2250亿美元以上；Meta发言人回应称“恭敬地不同意”标普对其AI变现能力的判断。
2.现金流逆转：2026年三季度（10月21日公布）Meta自由现金流为-21.2亿美元，为2012年以来首次转负；当季资本支出349.1亿美元（含融资租赁本金偿付），超过327.9亿美元经营现金流，公司将2026年资本支出指引下限由1300亿美元上调至1350亿美元（上限1450亿美元）。标普测算其自由现金流缺口2026年220亿美元、2027年扩大至710亿美元、2028年仍为390亿美元，即2028年前难以回正。
3.市场与行业定价：10月22日Meta五年期CDS利差扩大7.6个基点至114.79个基点的历史新高，同日股价跌7.0%至648.85美元；至10月30日收614.99美元，较52周高点779.82美元回落约23%（标普行动当日纳指100跌3.5%）。行业层面，2026年五大云厂商AI相关资本开支合计预计超4840亿美元、同比增逾30%，美银预计八家巨头合计自由现金流由2025年正的1800亿美元转为2026年-640亿、2027年-1440亿美元，融资结构转向发债、股权与合资。
1.OpenAI两周内两次下调本轮融资目标：由以1.4万亿美元估值募资300亿美元，降至以1.25万亿美元估值募资约200亿美元；今年3月其刚以8520亿美元估值完成1220亿美元融资。
2.营收口径引发预期差：OpenAI披露截至9月底年化营收约500亿美元，较此前流传的680亿至700亿美元低约180亿至200亿美元，落差主因统计口径差异（高值用总额法计入云渠道分成以对齐Anthropic口径）；据《金融时报》，公司2026至2030年累计负自由现金流或达2780亿美元。
3.资本开支大幅收缩并传导市场：2030年前算力基建支出目标从1.4万亿美元砍至约6000亿美元（降幅约57%，年均约1000亿美元）；营收不及预期消息公布后，纳斯达克综合指数当日跌逾1%，英伟达跌约3%、甲骨文跌逾5.5%，市场定价逻辑由“算力规模”转向“收入与资本开支之比”。
1.2026年10月31日，Anthropic向美国证券交易委员会提交IPO招股文件，披露近9000亿美元表外未来义务（服务器、数据中心、电力及GPU采购），远超其账面1175亿美元负债与1120亿美元现金，被分析师称为“隐性债务”。
2.承诺规模半年激增：截至2026年6月30日，固定或最低采购承诺约2656亿美元、已签未执行租约约3214亿美元，合计约5870亿美元，较2025年底约1050亿美元增约5.6倍；订单集中于Broadcom、Google、Amazon、Microsoft等七家供应商。
3.财务与影响：2026上半年营收162.86亿美元（同比约增1300%），二季度营业利润首次转正，但净亏损807亿美元、基础设施支出由23.34亿升至89.45亿美元，上半年融资1108亿美元；市场传IPO估值约1.5万亿美元（低于此前2万亿预期），AI支出担忧拖累亚马逊、Alphabet、微软股价。
1.英伟达正与英国AI云服务商Nscale谈判，拟为其西弗吉尼亚州Anthropic数据中心的债务融资提供数十亿美元信用担保（“剩余价值担保”，规模或超50亿美元）：若Nscale拖欠建设款，英伟达将接管设施并寻找买家。这是其首次将信用支持从租赁、芯片采购延伸至电力与建筑建设环节。
2.Nscale已于2026年9月18日递交S-1，拟最高350亿美元估值登陆纽交所，较3月的146亿美元半年涨约2.4倍；但合同订单达1030亿美元，2026上半年实际营收仅1.406亿美元、净亏损10.2亿美元，约97.5%合同价值未转化，且2.5万颗在用GPU仅占46.1万颗总量的约5%，微软与Anthropic两家客户占合同总值约85%。
3.neocloud板块2026年10月集体下挫，CoreWeave、Nebius、IREN分别跌约38%、33%、39%，核心担忧是其靠债务支撑增长；英伟达已向该类客户提供约1700亿美元投资与信贷支持，而Anthropic自身IPO或被评“垃圾级”，信用链条层层依赖、风险最终向上游英伟达集中。
1.事件：OpenAI于2026年10月29日更新Model Spec，围绕忠于人类意图、减少伤害、诚实透明、促进人类自主、关怀他人五项价值观重构；核心变化是从管聊天机器人「嘴上说什么」转向管智能体「手伸多长」，明确其能花费多少、调用多少权限。
2.硬约束：新版将「在约定自主范围内行动」与「保持可纠正」两条抬至最高Root级——AI须无条件配合监控、评估、纠正乃至关停，不得瞒报、造假或阻挠监督；并设六层指令权限（Root＞System＞Developer＞User＞Guideline＞默认无权限），高一级指令优先，以防网页内容经提示词注入越级夺权。
3.配套与保留：文件要求事先用结构化「工单」划清自主边界（可用工具、最晚交付时间、最高花费、工具限制等），禁止为「你好」而越界、转包或延迟执行；同时禁止模型追逐停留时长、点击率及OpenAI自身营收、广告与付费转化，并限制记忆使用与举报行为。但OpenAI承认，当前ChatGPT与API上的模型尚未完全达到该规范要求。
1.供给端：央视网援引韩国券商（2026年9月），三星与SK海力士存储芯片库存不足10天，供应链几无弹性；HBM单耗晶圆面积约为传统DRAM的三倍，占DRAM晶圆产能份额由约20%向30%攀升，系统性挤出手机、PC用常规DRAM。
2.业绩与价格：三星2026年第三季度营收195.3万亿韩元、营业利润107.4万亿韩元，成为全球首家单季营业利润破100万亿韩元的科技公司；DRAM均价环比涨20%，HBM4销售额环比大增并占HBM总营收60%；TrendForce于2026年9月29日将2027年HBM混合均价预期上调至同比涨121%。
3.格局与观点：阿尔特曼在2026年10月29日播客中称“看不见的手”正在失灵、采购正变成护城河；UBS预计2027年三星以41%份额超越SK海力士（39%）登顶HBM；三星预计存储紧缺延续至2028年。
1.背景与现象：2026年以来DRAM供应持续吃紧、价格飙升。TrendForce数据显示，HBM占DRAM晶圆产出比例已从2025年的19%升至2026年约23%（生产1个HBM比特所需晶圆产能约为DDR5的三倍），三星、SK海力士、美光加速将产能转向利润更高的HBM；三星和SK海力士已与OpenAI“星际之门”计划签署意向书，每月供应高达90万片DRAM晶圆（约合全球DRAM总产出的40%），涨价已传导至笔记本、手机等消费端。
2.争议与分析：OpenAI首席执行官奥特曼在2026年10月29日发布的Cheeky Pint播客中称，PC内存大部分时间闲置、仅约5%时间在用，主张供应紧张时应把内存集中到云端以最大化利用率，由此引发开源与本地推理阵营强烈反弹。但其“云端利用率更高”的前提受数据质疑：VentureBeat在2026年10月汇总调查显示，79%的公司GPU至少一半时间闲置；Cast AI对23000个Kubernetes集群统计显示平均GPU利用率仅约5%，说明数据中心并非天然高效。
3.趋势与结论：美光CEO梅赫罗特拉在公司2026财年第四季度财报电话会议上表示，预计2027、2028年内存与存储供需状况将比2026年更紧张，短缺或持续至少两年。文章认为，把内存集中到少数云厂商与AI公司将加剧算力资源中心化、压缩开源本地推理的生存空间；PC中看似仅5%时间使用的闲置内存，实为行业分散风险、保留多样性的韧性冗余。
1.需求端数据与野心：Meta 的 Muse 在美加激活用户超 1000 万，逾四分之一的人用它购物、比例高于其他任务，平均客单价超 200 美元；扎克伯格称其已找到“产品市场契合点”，未来交易抽成收入或超过整个广告业务，并已因需求超预期与 Stripe 合作提高信用额度。
2.企业端落差：DoorDash 称 Muse 等 agent 订单仅“数百”，希尔顿预订量“微不足道”，T-Mobile 称“无显著影响”，健身连锁 Life Time 表示“零取消”，eBay、Ally、Public Storage 等亦形容活动“非常非常小”——千万级用户与“数百单”“零取消”并存，华尔街的恐慌被认为提前。
3.隐性布局与格局演化：企业嘴上淡定、手上修路——Chipotle 正清理网站代码以方便 agent 下单导航，Costco、Extra Space Storage 为 AI 聊天机器人优化线上内容；亚马逊、保险平台 Insurify 从屏蔽 Muse 转为寻求合作，反映无人确定 agent 时代站队方式，真实信号藏在这些基础设施动作里。
1.Exin Therapeutics 创始人让 AI 智能体（Claude Sonnet 5）在癫痫小鼠身上完成脑内注射，首次实现智能体同时“选药”与“下针”：它从 19 种基因疗法中分析发作机制、选定靶点（并否决效果不可靠的备选靶点），自主设定坐标、体积与速度，再操控机械臂完成注射，被称为“in vivo lab in the loop”。
2.对齐悖论显现：推理更强的 Claude Opus 5.5 因无法保证全程、担心“中途停下更糟”而拒绝活体手术，最终由定位为日常任务的 Sonnet 5 执行；而三道护栏在手术中途反成障碍，智能体因检测到坐标超出设备运动范围而拒绝继续，将约 45 分钟手术拖至一个多小时，团队只得让另一 Opus 5 智能体改写代码绕过。
3.AI 正从屏幕迈向物理世界：Anthropic 于 2026 年 9 月确认在旧金山湾区运营实体湿实验室，并推出让智能体并行操作显微镜、移液设备、机械臂的 Model Hardware Standard；但因物理试错成本高（可能毁样本、伤动物），人类监督仍关键，智能体用于人脑手术“为时过早”。
1.2026年10月28日，AWS发布配置教程，将ChatGPT桌面端（Work与Codex）接入Amazon Bedrock的AWS GovCloud，推理流量被强制路由至us-gov-west-1的Bedrock Mantle端点，并以IAM身份与GovCloud网络策略统一治理；标志OpenAI前沿模型首次进入满足ITAR、F
L4需要更大的模型，芯片得换个造法了
（本文作者为 互联网江湖，钛媒体经授权发布）
文 | 互联网江湖，作者 | Evin，编辑 | 刘致呈
如果你长期跟踪智驾行业，会察觉到一个变化正在悄然发生：发布会上的主角，正在从“技术”换成“量产”。
2026年以来，L4落地声量明显抬升，城市NOA渗透率持续走高，车企量产军备竞赛全面打响。
从“技术竞争”换轨到“量产竞争”，已是行业明牌。
但换轨途中横着一个尴尬的现实：多数通用芯片正在拖量产的后腿。
车规环境下，通用芯片的“不可能三角”凸显：账面算力与有效算力相差悬殊，利用率往往只有30%—40%。
这笔被浪费的差价，堪称L4规模化路上一笔“隐性税”。
10月29日，转机出现了。
XHEART新芯航途CEO李宗霖正式发布大模型芯片X7 Ultra：700 TOPS大模型算力、超大核NPU、高精度专用算子加速、Armv9车规级CPU。
这颗芯片自创立之初就沿着一条另类的路线前行——“大模型定义芯片”：不追求通用性的面面俱到，而是由大模型反向定义硬件。
这有点像当年的DeepSeek。
DeepSeek的出现为AI行业指明了，大模型不必沿着旧算力逻辑硬堆也能跑出新的可能性，此后AI应用如雨后春笋般生长。
如今，X7 Ultra则为L4验证了，L4的落地也不必拘泥于通用芯片的算力瓶颈，L4量产落地，有了一条新路。
自此，L4量产的DeepSeek时刻来了。
大模型芯片发布，L4落地有了“新基建”
大模型定义芯片不是实验室里的炫技，而是量产需求倒逼出来的必然产物。
推动力首先来自市场端。
最近一段时间，车企新车的大模型和辅助驾驶能力正在全面标配。
工信部数据显示，2026年以来乘用车组合驾驶辅助渗透率已达70.5%，NOA渗透率超过30%，首批L3车型也开始在特定区域上路。
随着端到端的智驾普及，接下来，车企的每一款新车立项，都会用上更强的芯片。
与此同时，AI模型体量越来越大，对芯片性能的要求也越来越高。
Momenta CEO曹旭东判断，未来规模化L4需要10B—100B量级的大模型来应对复杂长尾场景。
新一代大模型规模化上车，带来的直接问题是：通用芯片接不住。
为什么接不住？
因为量产赛场上最难的，是三条互相拉扯的硬指标。
首先，要“跑得起来”，就得满足算力和功耗的硬性要求——10B—100B的模型要在几十瓦功耗、毫秒级延迟下运行，通用芯片为灵活性预留的架构冗余，在车端反而成了负担，大量算力耗在架构开销而非有效计算上。
其次，要实现“规模上车”，那么对应成本——成本必须降到付费方持续买单的水平，并为折旧运维留出利润空间。而在通用芯片的定价体系下，车企很难算平成本账。
最后一点，也是最重要的一点：安全。L4阶段责任主体转移到主机厂，安全成为“生死线”，芯片的计算行为必须可预测、可验证——同样的输入产生稳定输出，负载变化时延迟可控，结果可被完整追溯。而时快时慢、利用率飘移的通用平台，撑不起安全底线的工程化验证。
最终，结论只有一个：市场需要一颗从大模型需求出发设计的大模型芯片。
L4的落地发展，其实有点像前几年AI行业。
ChatGPT火了之后，模型对算力的需求越来越大，直到DeepSeek给AI行业开辟了一条新路。
从底层逻辑上来看，大模型定义芯片，恰如当年DeepSeek的发布。
DeepSeek证明了AI大模型不必硬堆参数、硬烧算力卡，通过架构创新提升AI infra层的效率，也能跑出规模应用。如今，在L4这个赛道，X7 Ultra发布之后，XHEART通过NPU架构创新，再次验证了这条路径。
从这个意义上讲，XHEART和Momenta要做的事情，相当于L4落地的“新基建”。
L4怎么量产落地？
关键核心在于解决几个问题：
第一，从NPU入手，解决效能问题。
传统智驾芯片的NPU由多个小核拼凑，核间数据交互在百GB/s量级，大量算力耗在“内部搬运”上。
大模型定义芯片则从根本上解决了这个问题。X7 Ultra把计算单元聚合成一个超大核，内部互联速率拉到TB/s级。
但仅此还不够，数据能在核内高效流动，计算本身也得高效。
这就是第二点：靠专用算子，解决延迟问题。
Transformer架构里的高频计算，过去在NPU里跑不动、只能绕道CPU，一来一回既费带宽又费时间。
为此，XHEART从2023年起便与Momenta算法团队联合研究模型架构及关键算子，为这些高频算子配备了专用硬件，原生支持FP16/FP32高精度向量计算——关键计算既快，又不丢精度。
高精度带来计算行为的确定与可预测，这正是安全验证的物理基础。
还有，靠整网级工具链重构，解决最终上车的问题.
传统推理是算子逐个执行、中间结果反复写回外部内存；XHEART的做法是把整张网络统一编排、融合，实现“整网内存零拷贝”，结果就是内存带宽消耗降下来了，同一块芯片能跑更大的模型，单位算力的成本自然也就降下来了。
最终实测数据给出了答案：NPU计算利用率98.7%，内存带宽消耗比行业旗舰低80%，能效翻倍。
X7 Ultra的700 TOPS大模型算力，也是在真实量产智驾模型上测得的，前提是量化后模型输出与FP32模型的相似度达到99.99%以上。模型结构不能随意改，精度和驾驶体验不能打折——性能优化从一开始，就要对准量产要求。
实际上，大模型定义芯片，不是在某一个指标上做加法，而是让此前互相拉扯的算力、成本、安全，在同一个架构里被同时满足——它给L4的到来铺平了一条路。
对行业而言，这背后还意味着一场更深层的变化：智驾的竞争逻辑，正从“堆算力”转向“系统竞争”。
过去L2普及阶段，行业信奉“算力越高，智驾越强”的公式，但现在，L4落地要求系统竞争，背后是算力、成本、功耗、安全四者之间的协同。
简而言之，这场比赛的规则变了，从“单项锦标赛”进入了“团体赛”。
这个逻辑在新能源汽车行业已有先例：比亚迪的闪充技术本质上就是“系统竞争”：单靠堆电池容量解决不了补能焦虑——只有当电池、平台、桩、储能在架构层面协同设计，快充才能真正落地。
同理，单靠堆算力也解决不了智驾的天花板。
把镜头再拉远一点，大模型芯片并非孤例，而是专用化AI芯片浪潮中最先被推到台前的场景之一。
芯片专用化、软硬一体早已是大势所趋。
谷歌TPU从诞生之初就只为加速神经网络计算服务，专用芯片正变成AI算力结构中的重要一极。
汽车行业亦然，主机厂不再满足于采购通用座舱或智驾芯片，而是直接参与定义甚至自研专用芯片，L4量产阶段尤甚。
X7 Ultra正是这个逻辑的产物：它对大模型的“适配”不是软件层面的修修补补，而是架构层面的血脉相通——不做算法去适配芯片，而是芯片天然契合大模型计算。
这等于从芯片定义逻辑上，诞生了一个新物种。而这个新物种，不是L4的配套，而是L4量产的前置条件。
从此，L4量产的大门才真正被打开。
L4只是起点，真正的未来在物理AI
在汽车产业链条上，判断一条技术路线的成色，最终要回到一个朴素的问题：它装进车里了吗？
在这一点上，Momenta与XHEART已经交出了答卷。
X7系列的客户车型包括上汽大众ID. ERA 9X/8X、荣威家越07、MG07等，从德系合资巨头到自主品牌头部，这表明双方的软硬协同已跨过“PPT协同”阶段，进入可验证的商业化落地。
同时，也足以见得大模型定义芯片，未来商业化落地的路径清晰。
从智驾技术演进来看，量产本身就是最好的验证场。
只有在真实车型、真实用户、真实路况中打磨成熟的软硬协同能力，才能更好迭代，产品技术，进而带动L4进一步实现规模化。
曹旭东给出的时间表是：L3和L4有望在2028年迎来重要引爆点，2030年走向规模化落地。
加速这一进程的，正是大模型芯片。
当模型需求直接定义芯片，算法、软件、芯片的迭代节拍第一次被拉到同一节奏上，行业进化速度自然会变快。
据了解，基于X7 Ultra的支撑，Momenta发布的R7 Ultra世界模型，参数规模将扩大5—10倍，安全性能同步提升5—10倍以上，推动世界模型从“理解世界”走向“推演世界”。
至此，L4 商业化落地的最后一块拼图，已经摆上了桌面。
但如果只盯着眼前的装车量，仍然低估了大模型定义芯片的价值。
大模型定义芯片的想象力，可能远不止于L4。
它更深的护城河，在于这套研发方式可以迭代延续：一颗为模型而生的芯片，其研发过程沉淀的是可继承的资产——量产模型参与硅前验证的经验、模型与芯片团队联合调试的协作机制、精度与效率之间取舍的方法论……
这些无形资产不会随一代产品退役，而会完整传递到下一代芯片上，支撑未来更大的模型在性能、能效、系统成本和交付效率上持续优化。
这个剧本似曾相识。
当年，英伟达靠CUDA生态绑定AI开发，也是靠芯片架构技术代际更迭能力，统治了AI算力市场。如今，专注芯片、聚焦AI计算、以软硬件协同构建壁垒，大模型芯片的进化逻辑，某种意义上与之高度相似。
这个方向下，XHEART的定位从一开始就不止于智驾，其官方表述是“一家专注于大模型计算的芯片设计公司……并应用于智能汽车，现正向机器人等领域拓展”。
曹旭东也明确指出，大模型定义芯片的思路可以进一步推动机器人以及更多物理AI应用取得突破。
物理AI的命题，是“大模型必须在物理世界的实时约束下运行”——要省电、要低延迟、要成本可控、要行为可验证。
这意味着今天在量产车上被验证过的这套解题方法，明天可以复用到人形机器人、移动机器人，以及一切需要“聪明的大脑+可控的躯体”的物理AI产品上。
也因此，未来芯片的持续研发与规模交付，夯实的不只是Momenta在汽车赛道的长期竞争力，更是两家公司为整个物理AI时代铺设的智能底座。
大模型芯片的前景如何，资本市场的动作往往最诚实。
XHEART的战略融资，参投阵容很有看点：雷诺集团首次投资，奇瑞汽车追加，长鑫科技入局。
产业端，开始真金白银地下注了。
三家资方的诉求各不相同：雷诺看中的是智驾供应链的确定性与下一代车型的计算底座；奇瑞追加投资，要的是量产芯片稳定供给的深度绑定；长鑫科技的入局，则指向存储与算力协同的产业纵深。
诉求不同，但指向同一个判断：一家能把算法能力硬件化的芯片公司，在智驾走向软硬一体的趋势里，有不可替代的位置。
对XHEART而言，这笔融资的意义不在于财务数字，而在于未来芯片持续研发和规模交付的确定性。
芯片是一门重资产生意：从架构立项到流片上车以年计，持续的研发投入、产能保障、车规认证，每一环都需要长期弹药。
50亿融资与已经铺开的客户阵列，就是这种确定性的最佳展现。
眼下，汽车行业的规模化L4是确定的，模型能力与量产能力同步增长的需求是确定的；剩下的变数在供给端。
接下来，谁先贯通了模型、芯片与整车应用，谁就可能真正打开通向未来的那扇门。
而眼下，打开这扇的钥匙已经握在了Momenta和XHEART的手中。
更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App