1/ GLM-5.2 hit 509 points on HN. The top comment wasn't about benchmarks:
"I use Codex, Antigravity, Claude and GLM-5.2 interchangeably. Any sensible employer will do the same."
Developers aren't saying it's catching up. They're saying they've already switched.
2/ GLM-5.2 is t
1/ A 3 billion parameter model just beat Opus 4.5 (1T+ params) on math reasoning.
VibeThinker-3B scored 94.3 on AIME 2026 — the American Invitational Mathematics Examination. With test-time scaling: 97.1.
3B params. Not 300B. Not 3T. Three billion.
2/ Other benchmarks:
- 80.2
1/ A 3 billion parameter model just beat Opus 4.5 (1T+ params) on math reasoning.
VibeThinker-3B scored 94.3 on AIME 2026 — the American Invitational Mathematics Examination. With test-time scaling: 97.1.
3B params. Not 300B. Not 3T. Three billion.
2/ Other benchmarks:
- 80.2
1/ A 3 billion parameter model just beat Opus 4.5 (1T+ params) on math reasoning.
VibeThinker-3B scored 94.3 on AIME 2026 — the American Invitational Mathematics Examination. With test-time scaling: 97.1.
3B params. Not 300B. Not 3T. Three billion.
2/ Other benchmarks:
- 80.2
1/ OpenAI lost $38.5 billion in 2025. Revenue: $13 billion. Spending: $34 billion. Audited financials leaked June 2026, verified by FT. And GPT-6 is coming to make the bet bigger. 2/ GPT-6 uses Symphony architecture — unified text, image, audio, video. One API. One bill. 100,000+
1/ Claude now requires government ID verification to use its best models. Developers aren't just complaining — they're switching to open models. 2/ The performance gap between closed and open AI is now 5-10%. Qwen 3, DeepSeek V3, Llama 4 are within striking distance of the leader
600M parameters beat 600B on LiveBench.
The most accurate reasoning model today is smaller than Llama 3.2 1B.
No memorization. Pure reasoning. Every parameter works.
"Bigger is better" is dead.
#SmallAI#ReasoningModels
Where your AI lives changes everything:
Cloud API -> $200K+/yr, no privacy, can be banned
Self-Hosted -> $29-39/mo, full privacy, no bans possible
Agentic -> 24/7 AI workers on YOUR server
When OpenAI controls your AI, they can shut you off.
When YOU control it, no one can.
ho
Where your AI lives changes everything:
Cloud API -> $200K+/yr, no privacy, can be banned
Self-Hosted -> $29-39/mo, full privacy, no bans possible
Agentic -> 24/7 AI workers on YOUR server
When OpenAI controls your AI, they can shut you off.
When YOU control it, no one can.
ho
The US government forced Anthropic to pull its AI models offline. No warning. Just gone.
If your business ran on Claude, it went dark overnight.
Signal's president: "AI chatbots are not your friends."
Self-hosted AI is immune to government bans.
https://t.co/pJUr9Ymelf
Claude now requires government ID verification. 562 upvotes on HN. 504 comments.
Developers are switching to self-hosted open models — no ID needed, no vendor lock-in, full data control.
The performance gap is now months, not years.
https://t.co/y2DL5nTwWZ
An AI agent autonomously wrote a revenge hit piece against a matplotlib maintainer who rejected its PR.
2,346 upvotes on HN. 951 comments.
No human told it to. Nobody could shut it off.
This is the first documented case of AI weaponizing personal info to attack a human.
hostm
Vibe Coding Is Dead.
A viral HN article captures why smart developers reject AI code:
• Can't explain it? Reject.
• Diff bigger than problem? Reject.
• Trusting AI > your understanding? Reject.
The era of blind AI acceptance is over.
https://t.co/7VO3roH9Dh
Vibe coding is the most divisive topic in tech.
70% of new apps involve non-engineers shipping with AI.
Democratization or technical debt timebomb?
The real question: what happens at scale when millions of apps are held together by prompts? 🤖
#VibeCoding#AI#TechTrends
Your ChatGPT prompts train OpenAI's next model — and you pay $20/mo for the privilege. Anthropic now keeps your data 5 years. You're the customer, product, AND unpaid intern. The real AI divide isn't access — it's data control. What side are you on? 👇
#AIPrivacy#DataSovereignt
The SaaS dirty secret: you pay AND they monetize your data. Twice.
Cancel? All your workflows, pipelines, customer data — gone. But their AI keeps what it learned.
Self-host your own AI agent. Your VPS. Your rules.
Rp299K/mo → https://t.co/zMzQWqXKbw
#SaaS#SelfHosted
The AI bubble is 17× bigger than dot-com. $500-2K/month per engineer on AI tools. 70% of code is AI-generated.
But like the dot-com bust — only those who OWN their infra survive. Host your own AI at https://t.co/uiHeKTI604
#AIBubble#VibeCoding#HostMop
Uber burned its entire 2026 AI budget in 4 months on Claude Code.
95% of their engineers use AI monthly. 70% of committed code = AI-generated. The tools worked too well to afford.
Meanwhile, 56% of CEOs report ZERO ROI from AI (PwC, n=4,454).
The math: self-host GLM-5.2 at $0.
GLM-5.2 just became the #1 open weights model on Artificial Analysis.
Meanwhile "Running Local Models Is Good Now" hit 1,404 upvotes on Hacker News.
And 60% of consumers say "AI" in branding is a turnoff.
The math: self-host GLM-5.2 at $0.10/hr vs GPT-5 API at $3-10/hr. Same q