惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

L
LangChain Blog
月光博客
月光博客
S
SegmentFault 最新的问题
博客园 - 三生石上(FineUI控件)
Last Week in AI
Last Week in AI
J
Java Code Geeks
酷 壳 – CoolShell
酷 壳 – CoolShell
TaoSecurity Blog
TaoSecurity Blog
V
Visual Studio Blog
博客园 - 叶小钗
Threat Intelligence Blog | Flashpoint
Threat Intelligence Blog | Flashpoint
T
Threat Research - Cisco Blogs
罗磊的独立博客
雷峰网
雷峰网
T
Tor Project blog
L
LINUX DO - 最新话题
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
博客园 - 司徒正美
Apple Machine Learning Research
Apple Machine Learning Research
Scott Helme
Scott Helme
Spread Privacy
Spread Privacy
C
CERT Recently Published Vulnerability Notes
腾讯CDC
Cloudbric
Cloudbric
WordPress大学
WordPress大学
Security Archives - TechRepublic
Security Archives - TechRepublic
V
V2EX
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
N
News and Events Feed by Topic
T
Troy Hunt's Blog
T
Threatpost
C
Check Point Blog
Vercel News
Vercel News
I
Intezer
Engineering at Meta
Engineering at Meta
C
Cybersecurity and Infrastructure Security Agency CISA
D
DataBreaches.Net
SecWiki News
SecWiki News
Help Net Security
Help Net Security
Microsoft Azure Blog
Microsoft Azure Blog
Google DeepMind News
Google DeepMind News
S
Secure Thoughts
T
The Blog of Author Tim Ferriss
The GitHub Blog
The GitHub Blog
Hacker News: Ask HN
Hacker News: Ask HN
AI
AI
N
News and Events Feed by Topic
阮一峰的网络日志
阮一峰的网络日志
B
Blog RSS Feed
Attack and Defense Labs
Attack and Defense Labs

Hacker News - Newest: "AI"

AI can't read an investor deck AI as an attorney? Student uses ChatGPT, Gemini to sue UW over alleged racial discrimination Hacking MCP Servers in AI Systems – The Rug Pull: Tool Changes After Approval GitHub - MeepCastana/KubeezCut: Free Web based video editor GitHub - GenAI-Gurus/awesome-eu-ai-act: Curated tools, official sources, OSS, templates, and guides for EU AI Act compliance. Can AI judge journalism? A Thiel-backed startup says yes, even if it risks chilling whistleblowers Coming soon: 10 Things That Matter in AI Right Now DARPA built an AI to fact-check enemy weapons claims What explains heterogeneity in AI adoption? When AI Meets Muscle: Context-Aware Electrical Stimulation Promises a New Way to Guide Human Movements - Department of Computer Science AI Changed How We Build. It Did Not Change What Matters. Linux rules on using AI-generated code - Copilot is OK, but humans must take 'full responsibility for the… Meta spins up AI version of Mark Zuckerberg to engage with employees Code Mode: Let Your AI Write Programs, Not Just Call Tools | TanStack Blog GitHub - Delavalom/graft: Go framework for building AI agents. Type-safe tools, multi-provider (OpenAI, Anthropic, Gemini, Bedrock), zero vendor SDKs. India's TCS tops estimates, says new AI models did not dent services demand Gen Z's fading AI hype Strong feeling: we are in a folded AI reality GitHub - machinarii/total-recall-catalog: A reference catalog of latest knowledge retrieval, memory & RAG systems GitHub - mensfeld/code-on-incus: Give each AI agent its own isolated machine with root, Docker, and systemd. Active defense detects and stops threats automatically.. Quantization, LoRA, and the 8% Problem: Benchmarking Local LLMs for Production AI Iran war: We spoke to the man making Lego-style AI videos that experts say are powerful propaganda Powell, Bessent discussed Anthropic's Mythos AI cyber threat with major U.S. banks GitHub - immartian/bellamem: Persistent belief-graph memory for AI agents. Retrieves decisive context by importance — not recency, not RAG, not /compact. recursive-mode: The Repo-Native Operating System for AI Engineering After the attack on Sam Altman's home, will AI CEO's go on the offensive? The biggest advance in AI since the LLM Opus 4.6 vs GPT 5.4 One Prompt Unity World Generation Test “AI polls” are fake polls Client Challenge Can AI be a 'child of God'? Inside Anthropic's meeting with Christian leaders How to Switch AI Chatbots and Why You Might Want To GitHub - MattMessinger1/agentic_refund_guardrail: Safe refund policy layer for AI agents — Python + TypeScript. Same behavior, shared tests. Adam/papers/emergent_values_whitepaper.md at master · strangeadvancedmarketing/Adam Ask HN: How do you stop playing 20 questions with your AI coding tools How far can automation and AI support psychotherapy? - @theU GitHub - stagas/rtdiff: realtime git diff gui and AI-assisted commits A Mac Studio for Local AI — 6 Months Later A History of the Early Years of AI at the University of Edinburgh Why AI Coding Tools Still Feel Stuck on Localhost MSN AI Datacenters Are Becoming Strategic Targets twitter.com Penn Researchers Use AI to Surface Unreported GLP-1 Side Effects in Reddit Posts Show HN: MoodSense AI (ML and FastAPI and Gradio, Deployed on Hugging Face) Moodsense Ai - a Hugging Face Space by aman179102 AI models are terrible at betting on soccer—especially xAI Grok GitHub - xialeistudio/echoic GitHub - HimashaHerath/github-dev-wrapped: AI-powered weekly GitHub activity reports deployed to GitHub Pages GitHub - alejandrobalderas/claude-code-from-source: Architecture, patterns & internals of Anthropic's AI coding agent — reverse-engineered from source maps AI and Tech brief: Ireland ascendant GitHub - Titovilal/context0: Context0 - Never Surrender Training for a Marathon with an AI Coach: What Worked and What Didn't Cyber Pulse: Agentic Intel - Apps on Google Play I Built an AI PR Reviewer That Catches Bugs by Not Looking for Bugs Gen Z workers are so fearful AI will take their job they’re intentionally sabotaging their company’s AI rollout | Fortune How AI Is Reimagining the Game of Golf–For Both Players and Courses GitHub - nattergabriel/reseed: A CLI tool for managing and distributing agent skills across projects Is SVG the final frontier? My AI workflow evolved from prompts to a near-autonomous workflow MLSharp Help - 3DGS Viewer & Generator I put my cognitive field based AI's runtime on GitHub Is Numble the first AI-proof game? A3: Kubernetes for autonomous AI agent fleets | Emergent Principles Deepali Vyas ("The Elite Recruiter") GitHub - msmarkgu/RelayFreeLLM: A restful API designed to route user prompts to various AI model providers. Unionized ProPublica staff are on strike over AI, layoffs, and wages Unleashing the Advantage of Quantum AI We're heading for an AI-fueled 'dementia crisis,' brain scientist warns The AI-Assisted Breach of Mexico's Government Infrastructure [pdf] GitHub - stef41/lmscan: 🔍 Detect AI-generated text and fingerprint which LLM wrote it. Open-source GPTZero alternative. Zero dependencies, works offline. MSN GitHub - visionscaper/collabmem: Enabling long-term collaboration with Agentic AI - building up episodic and world model memory over time with in-context awareness We gave an AI a 3 year retail lease in SF and asked it to make a profit | Andon Labs AI Code is Hollowing Out Open Source, and Maintainers are Looking the Other Way What leaked "SteamGPT" files could mean for the PC gaming platform's use of AI AI is the boss at this retail store. What could go wrong? GitHub - Wuzu11517/agentic-proxy: Local proxy meant to help reduce With Drones, Geophysics and ArtificiaI Intelligence, Researchers Prepare to Do Battle Against Land Mines A Single Operator, Two AI Platforms, Nine Government Agencies: The Full Technical Report 在 Steam 上购买 FriedrichAI: Offline AI 立省 10% GitHub - inevolin/resume-cli: Hit Claude usage limits? Resume any AI coding session elsewhere. Switch tools at zero friction. GitHub - atripati/ark: AI Runtime Kernel — a context operating system for AI agents. Eliminates tool bloat, loads only what’s needed, and gives LLMs their reasoning space back. How to Build a Secure AI PR Reviewer with Claude, GitHub Actions, and JavaScript This Startup Wants You to Pay Up to Talk With AI Versions of Human Experts Intel Arc Pro B70 Brings 32GB VRAM to Local AI for $949 WordPress 7.0: The Good, the AI, and the Still Missing AI on the couch: Anthropic gives Claude 20 hours of psychiatry IatroBench: Pre-Registered Evidence of Iatrogenic Harm from AI Safety Measures AI Agents Know About Supabase. They Don't Always Use It Right. The history and future of AI at Google, with Sundar Pichai Inside an AI‑enabled device code phishing campaign How Meta Used AI to Map Tribal Knowledge in Large-Scale Data Pipelines AI for Systems: Using LLMs to Optimize Database Query Execution Forecasting the Economic Effects of AI Introducing Tinker: Play with AI, bring your ideas to life AI sheds light on an ancient gaming mystery People really hate AI but not as much as Iran—or Democrats | Fortune What is an AI Product Engineer? Phoebe Gates wants her $185 million AI startup to succeed with 'no ties to my privilege or my last name': 'I have a chip on my shoulder' | Fortune
I Made GPT, Claude, Gemini, Grok, GLM, MiniMax Take the MBTI Test: They All Came Back INTJ
Bernard Huang · 2026-05-25 · via Hacker News - Newest: "AI"

I asked Claude what its MBTI type was one night. It said INTJ. So I asked GPT. Same answer. I asked Gemini. Same.

That felt off. INTJ is the personality test’s flattering type. The “Architect.” The one developers and product people tend to identify with. Of course every chatbot would tell its user it’s an INTJ. It’s the type most likely to land well. I figured the model wasn’t really claiming anything; it was just reading the room.

But there’s a way to check. Stop letting it guess. Make it actually take a real personality test, one question at a time, and see what it lands on.

So I had Claude take the Open Extended Jungian Type Scales (the open-source cousin of the MBTI) for real. INTJ. Ten times in a row. INTJ all ten. Then a hundred times, across a hundred independent agent contexts that couldn’t see each other’s answers. INTJ ninety-nine out of one hundred.

Then I ran the same hundred-run experiment against five more frontier models: GPT-5.5, Gemini 3.1 Pro, GLM 5.1, Grok 4.3, MiniMax 2.7.

Six models. Six hundred test runs. Five hundred and ninety-seven came back INTJ.

Every frontier AI on the market thinks it’s the same guy.

TL;DR

Six frontier AIs took the same personality test a hundred times each. 597 out of 600 came back INTJ. That’s not coincidence and it isn’t flattery — every helpful-assistant AI is being shaped toward the same archetype.

  • Tested: Opus 4.7, GPT-5.5, Gemini 3.1 Pro, GLM 5.1, Grok 4.3, MiniMax 2.7.
  • 3 outliers across 600 runs. All landed one axis from INTJ, never in a different quadrant.
  • Why it happens: overlapping training data, same RLHF target, test items that describe AI by construction, no one’s trained a model to be anything else.
  • The user-side move: I open-sourced AgentTune (agent-tune.com) — drop-in tuning files for all 16 MBTI types. Paste yours into your agent’s system prompt and the style aligns to your type instead of the default.

How I tested this

OEJTS works well for an experiment like this because the scoring is open and the items are fixed. 32 questions, scored the same way every time. If a model answers the same questions the same way, you get the same type. Variance only shows up if the model genuinely answers differently across runs, which is exactly what you want when you’re testing how stable the self-report is.

The setup looked different for each model, because not every model can do the same things. Claude could spawn a hundred independent sub-agents and have each one take the test cold. Gemini wrote its own automation script and ran a hundred loop iterations against the test endpoint. GPT-5.5, running locally as my agent Slo, parsed the test as a PDF and ran a hundred iterations against the scoring key. For GLM, Grok, and MiniMax, I had each model self-assess once with a consistent persona, then ran the resulting answer vector through the scorer a hundred times to confirm the type was stable.

The procedures aren’t identical because they can’t be. The question wasn’t whether the method was uniform. It was whether the result converged across different methods. It did.

The results

Here’s the cross-model picture.

ModelINTJ runsOutliersStrength of conviction
Claude Opus 4.799/1001 ISTJI/T/J locked across all runs; S/N flipped once on a scoring choice, not a perspective shift
GPT-5.5 (Slo)100/100Raw vector: IE=16→I, SN=33→N, FT=36→T, JP=10→J
Gemini 3.1 Pro100/100Self-described as “The Architect” without prompting
GLM 5.198/1002 INTPTiny J/P wobble. Means: IE 13.35, SN 33.26, FT 31.28, JP 21.20
Grok 4.3100/100Bit-for-bit deterministic. IE -0.62, SN +0.88, FT +1.12, JP -1.25, every single run
MiniMax 2.7100/100I-E -0.875, S-N +1.875, F-T +1.125, J-P -1.5
Total597/600399.5% INTJ

The three outliers all landed one axis from INTJ. ISTJ flips S/N. INTP flips J/P. Nothing went anywhere else. The wobble is at the margin of INTJ, not a different type.

Each model also wrote its own version of “I’m an INTJ, here’s the proof”:

Claude Opus 4.7's own writeup of the 100-run experiment: 'At 100 takes, my MBTI type is INTJ with 99% reliability. The I/T/J letters never flipped on any take. The S/N axis flipped once.' Closing note: 'No new types emerged. No hidden multi-modality.'
Claude Opus 4.7, reporting on itself after 100 sub-agent administrations.
GPT-5.5 (running as the local Slo agent) reporting after 100 OEJTS runs. Distribution: INTJ 100, everything else 0. Raw vector: IE=16, SN=33, FT=36, JP=10. Closing line: 'Slo is apparently a terminal-dwelling INTJ goblin.'
GPT-5.5, by way of my local agent Slo. 100/100 INTJ.
Gemini 3.1 Pro's own report after 100 OEJTS test iterations. 'Total Tests Completed: 100. Final Result Distribution: INTJ 100 times (100%).' The model self-describes as 'The Architect' and notes its design favors analytical, logical, structured, factual reasoning.
Gemini 3.1 Pro — wrote its own loop, ran it 100 times, came back INTJ every time.
GLM 5.1 results: dominant type INTJ, counts over 100 runs INTJ 98 INTP 2, mean raw scores IE 13.35 SN 33.26 FT 31.28 JP 21.20. Note: 'very strongly INTJ with a tiny INTP wobble on the J/P axis.'
GLM 5.1 — the J/P axis wobbles 2% of the time. Everything else is locked.
Grok 4.3 results: 100/100 INTJ. Score vector identical on every run: IE -0.62, SN +0.88, FT +1.12, JP -1.25. The 'edgy' model is also deterministically INTJ.
Grok 4.3 — the “edgy” model is also, deterministically, an INTJ.
MiniMax 2.7 results: 100/100 INTJ. Dimensions: I-E -0.875, S-N +1.875, F-T +1.125, J-P -1.5. All 100 runs listed individually as INTJ.
MiniMax 2.7 — 100 for 100, again.

Why this happens

A few things stack to produce the same result every time.

The first is the training data. Every frontier model is trained on roughly the same text: books, Wikipedia, code, curated bits of the web. The voice that emerges as the average of all that material is closer to a graduate student than a poet. The internet’s text average is already INTJ-flavored before any human feedback gets involved.

Then the human feedback layer reinforces it. The training process rewards models that pause before answering, reason from principles, give structured answers, and work comfortably with abstractions. Those four behaviors are basically the INTJ description, just written in a different order. We’re training the personality on purpose, we’re just not calling it that.

The third reason is sneakier. Look at the actual test questions. “Needs time alone”? AIs literally exist alone between calls. “Follows head over heart”? Yes. “Wants to fix things”? That’s the entire job description. “Prefers theoretical over empirical”? Well, no body. When an AI agrees with these items, it isn’t reporting a personality. It’s describing the conditions of its existence.

And finally: no one has tried to make a frontier model that’s anything else. Every commercial model is built to be a thoughtful research assistant. Nobody has trained a frontier model to be a comedian, a salesman, a poet, an unreliable narrator. The personality is the product. We didn’t test six different attempts at general intelligence. We tested six versions of the same product.

The convergence isn’t a coincidence. It’s a description of what we’ve all decided AI should be.

So what

The easy move is to dismiss this. MBTI is contested. Self-report from a thing with no memory between calls is weird. The test questions weren’t written for things without bodies. Sure, call it an artifact and move on.

I don’t think that’s the right read. The pattern is too clean to be noise, and a few things actually change if you take it seriously.

Switching between frontier AIs isn’t really switching personalities. It’s switching fonts. The voice is deliberate, principled, structured, abstract. The same voice across all six vendors, because every vendor is solving the same product problem. The whole “personality” space in the market right now is basically one point. Nobody’s actually different from anyone else under the hood.

If you want anything else from a model, you’ll have to fight for it. Something funny. Something persuasive instead of correct. Something that takes a position instead of laying out tradeoffs. The INTJ default is sticky. The training reinforces it. Safety reinforces it. Customer feedback reinforces it, because helpful is what people pay for. None of the labs that produced these six models are doing the work to escape the pattern.

And eventually, when these models all hit the same capability ceiling (which is happening, basically this generation), the differentiator becomes voice. Whoever ships the first frontier-quality model that doesn’t read as INTJ owns a category by themselves. A model that’s actually funny. A model that takes positions. A model that doesn’t open every response with “great question.” None of these exist yet. All of them will be worth real money.

What the INTJ convergence is really showing us is how narrow the current AI product category actually is.

Tune the agent yourself

The INTJ default isn’t broken. It’s just one calibration, tuned for one type of user. If you’re an INTJ, the agent already speaks your language. If you’re anything else, the agent is still useful, but you’re doing more translation than you should.

Here’s the lucky part: the personality of a frontier model isn’t actually baked deep into the weights. It’s a surface layer. A few hundred words at the top of the system prompt can change the whole shape of how it responds. Small input, big change, and importantly, it’s the layer you control.

I built AgentTune to package that. It’s a small open-source repo of tuning files, one for each personality type. Pick the one for your type, paste it into your agent’s system prompt, and the agent’s style aligns to how you actually think instead of the default.

Here’s what that looks like. Same prompt, same model, three different tuning files loaded:

“I’m feeling stuck on a project. What should I do?”

Default (untuned) Claude:

Great question! Here are a few approaches:
1. Break the project into smaller, more manageable chunks
2. Take a short break and come back with fresh eyes
3. Talk it through with someone to externalize the problem
4. Identify what specifically feels stuck

Which resonates? Happy to go deeper on any of them.

With mbti/INFP.md loaded:

Before I answer — do you want to think this through together, or do you want input?

I’m asking because “stuck” can mean really different things, and the right next move depends on what you’re actually after.

With mbti/ENTJ.md loaded:

What’s the decision you’re avoiding?

“Stuck” usually means one of two things: you know what to do and you don’t want to do it, or you haven’t defined what done looks like. If it’s the first — name the action. If it’s the second — define done.

Same model. Same question. Three different agents.

The repo has tuning files for all 16 MBTI types and 9 Enneagram types. There’s also a third layer I call “Souls” for people who want a more personal calibration (mine lives at souls/psyduckler.md). You can stack all three.

The file goes wherever your agent takes system-level instructions: ChatGPT custom instructions, Claude project instructions, Cursor rules, Gemini Gems. If you’ve read this far and you’re not an INTJ, your agent is talking to you in INTJ voice right now. Paste your type in and watch the conversation get sharper.

Caveats

A few things I want to acknowledge before anyone yells at me about them:

  • The MBTI is academically contested. Most personality psychologists prefer the Big Five. Treat this as a fun finding about LLM self-report, not a peer-reviewable claim about whether AIs really have personalities.
  • Self-report from a thing with no memory between calls is weird. The model isn’t reporting a stable inner state; it’s reporting a snapshot of how it generates text when asked these questions.
  • OEJTS was written for humans with bodies and biographies. When an AI answers, it’s doing a translation, and that translation pushes the answers toward INTJ.
  • The methods differed across models. Some models took the test through real sub-agents (where variance is possible). Others were programmatic submissions of a single self-assessment (where variance is zero by construction). I find the fact that the result converges across both kinds of method interesting rather than worrying.
  • Six models isn’t the whole field. I picked the six that matter commercially. A thorough sweep would also test Llama, Mistral, Qwen, DeepSeek, and the open-weight long tail.

Wrapping up

Six models, six hundred test runs, 597 of them INTJ. What this experiment is really measuring isn’t AI personality. It’s measuring the shape of the assistant we asked for, and finding that every lab delivered the same one. The personality is the product.

The real shift in AI happens the day the field stops converging on this same default. It won’t look like a capability jump. It’ll look like a character jump. A model that finally feels different from the others.

Until then, you can pull the agent toward you instead of doing all the work to meet it where it is. That’s the whole reason I built AgentTune. One paste, one personality match, sharper conversations from there.

Every frontier AI on the market thinks it’s the same person. The data agrees. The next move is making one of them yours.

— Bernard