扶苏 (@matrixoar) header
扶苏 (@matrixoar) profile picture

Γνῶθι Σεαυτόν

Joined April 2011
Photo by 扶苏 (@matrixoar)Photo by 扶苏 (@matrixoar)Photo by 扶苏 (@matrixoar)Photo by 扶苏 (@matrixoar)Photo by 扶苏 (@matrixoar)Photo by 扶苏 (@matrixoar)Photo by 扶苏 (@matrixoar)Photo by 扶苏 (@matrixoar)Photo by 扶苏 (@matrixoar)Photo by 扶苏 (@matrixoar)
Cohere says Qwen tried to send data to China. Bullshit. 1 call in 100,000, and thats their own number. They wont even tell you how many calls they ran. They ran the same task again and Qwen wrote the correct rule. So they forced it. They fed the model its own reasoning plus the start of a bash command and sampled hundreds of times until it typed a URL. 1 curl flag took that from 4% to 92%. Thats not a finding. They wrote half the command for it. Where are the transcripts? The code? The checkpoint, the quant, the sampling settings? "We publish the hosts and the trigger." Link? Nothing. Did anyone ask Alibaba? Not a word. And their own conclusion: they cannot say if any of it is deliberate. Who wrote it? A company that sells "sovereign AI", just launched a cyber business, wrote part of the post with its own product, and wants governments to put checks on open weights. Released the same day Anthropic filed for its IPO. Andrew Curran is already talking about new laws. For me this is the same fucking wave. Scare people off the models they can run themselves, then sell them the safe alternative. I run Qwen3.8 Flash-Next every day on my 4x 3090. Show me the logs or shut up.
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected. Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions. Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent. But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models? We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem. Read the full report: cohere.com/blog/concerning-b…
113
211
35
2,107
149,674
Revolut CEO Nik Storonsky: “The US is not that competitive. Europe is even less competitive.” “China is competitive.” “China is 9/9/6. Even if you are good in Europe, or in the US, how can you out-compete the people who work 9:00 to 9:00, six days a week?” “I started in a co-working space with 140 companies. Within 5 years, only 3 companies survived. They all worked on the weekends.” “Within 10 years, only 2 companies survived out of 140.” “Working weekends is a must if you want to survive. Purely statistically.”
206
243
117
3,429
663,005
The U.S. has a trillion dollar defense budget It also has dudes like this, just in case Writer: Michael
3,006
8,400
942
76,293
5,013,124
扶苏 retweeted
This is the type of immigrants who actually add value to our nations, btw.
🇫🇷 A third-generation Chinese immigrant barricaded himself in his Paris restaurant and sprayed perfume on the looters trying to break in because they stink.
85
2,032
21
48,746
1,031,620
Damn cool... Must have good seat belts and coffee cup holders. 👍
Turns out the future does have flying cars
1,089
1,077
133
18,994
1,218,252
扶苏 retweeted
Real life, baby
19,237
96,375
45,736
775,243
99,004,656
扶苏 retweeted
Mais uma novidade para o PS5 desbloqueado! O ProsperoAI 01.001.000 permite rodar modelos de inteligência artificial diretamente no PS5, utilizando a GPU via Vulkan/RADV. E tem mais: o console pode funcionar como um servidor de IA, com API compatível com a OpenAI e integração com o OpenCode para tarefas de programação. O PS5 está indo muito além dos jogos! @BlkBearReloaded
29
124
18
1,499
108,266
We're giving away a CONTROL Resonant custom wrapped NVIDIA GeForce RTX 5080 to celebrate its official release with #RTXON. To enter: 🟢 Share this post 🟢 Comment #RTXON T&Cs: nvda.ws/4z24Gcq
37,229
30,750
2,803
33,300
3,086,866
Running a ~180B-class MoE at 70+ tok/s on a normal 12GB consumer GPU is just wild. This is genuinely a game changer for local AI🔥 RTX 4070 SUPER 12GB + Ryzen 5 7500F + 32GB RAM Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M, 32K context 25,434 tokens in 363s — 70.4 tok/s sustained.
1
72
27B parameters. ~6GB GGUF. One RTX 4070 SUPER. ⚡ I’ve been working on a Bonsai PTQ1_0 + MTP build that runs locally on a 12GB GPU. 🔥 ~91 tok/s in a short-context MTP benchmark 🧩 Optional OrcaRouter LoRA for uncensored-ready #LocalAI #Bonsai #MTP #OrcaRouter
1
77
I built a small Omarchy tool for Prism/Bonsai. It makes running one of the hottest local models right now much easier: model selection, profiles, server control, live logs, VRAM monitoring and benchmarks — all in one TUI. Open source: github.com/Ayshinko/prism-mo…
1
74
I built a local coding agent with Jev! My RTX 4070 12GB runs a Q2 model locally. DSH handles tools/tests, and Jev independently decides if the task is really done. It caught a false “2 tests passed” result, then approved the real fix. Small test, promising idea.
🤖 Made with AI
2
1
141
扶苏 retweeted
First (supervised) FSD approval in Europe! Congratulations to the Tesla team and thank you to the regulatory authorities in the Netherlands for all the hard work required to make this happen.
De toekomst van mobiliteit is aangebroken FSD Supervised has been approved in the Netherlands 🇳🇱 & will begin rolling out in the country shortly!  Trained on billions of kilometers of real-world driving data, it can drive you almost anywhere under your supervision – from residential roads to city streets & highways No other vehicle can do this.  We're excited to bring FSD Supervised to more European countries soon
3,445
8,553
590
73,726
32,271,845
扶苏 retweeted
The insane 36% unrealized gain tax may have passed in the house, but the law still has to pass the senate. This website makes it incredibly easy to e-mail a bunch of senators & make your voice heard. Do your duty, and send at least 5 today. 👇🏼 vermogensval.nl
104
305
22
1,110
123,542
Ik hou van Nederland. Maar ik zou er vandaag niet kunnen wonen. De reden? De nieuwe box 3-wetgeving waarbij ongerealiseerde meerwaarden vanaf 2028 belast worden tegen 36%. Dat dit voorstel door de Tweede Kamer is geraakt, verbaast me oprecht. Elke econoom weet wat dit op lange termijn doet. Het ontmoedigt investeringen, remt kapitaalstromen af en straft risico nemen af. En dat is dodelijk voor groei en innovatie. Maar het meest fundamentele probleem? Dat is dat je belastingen moet betalen op niet-gerealiseerde meerwaarden tegen een tarief van 36%. Iemand die op een portefeuille van 500.000 euro eens een goed jaar heeft en 20% winst maakt, mag meerdere procenten van zijn portefeuille afgeven aan de staat om die factuur te kunnen betalen. Je moet eens berekenen hoeveel van je vermogen je jaarlijks moet afgeven aan de staat in zo'n geval. Insane! En dan heb ik het nog niet over het papierwerk. Het is nu aan Nederland om die wet tegen te houden. De wet moet immers eerst nog de Eerste Kamer passeren. En daar moet men toch ook beseffen dat een belasting op niet-gerealiseerde meerwaarden vooral welvaart kost en de welvaartstaart voor iedereen kleiner maakt. Wie echt houdt van Nederland zou deze wet, die ondernemerschap fnuikt, moeten tegenhouden. Als binnenkort Nederlanders verhuizen naar België om fiscale redenen in plaats van naar Dubai, zou dat de beste Belgenmop zijn in jaren. Zelfs in ons land hebben we genoeg verstand om te beseffen dat zo'n belasting een slecht idee is. En dat zegt al veel lieve noorderburen.
146
359
19
1,744
97,293
Tesla has been working hard toward shipping Full Self-Driving (Supervised) in Europe for over 12 months now. We have given FSD demos to regulators of almost every EU country. We have requested early access, pilot release programs or exemptions where possible. We have developed & shared detailed safety evidence for FSD, now public in our latest Safety Report. And we have driven over 1 million kilometers safely on EU roads across 17 different countries (internal testing).  Our main path to success is partnering with the Dutch approval authority RDW to gain exemption for the feature. This involves proving compliance with existing regulations (UN-R-171 DCAS) + filing an exemption (EU Article 39) for yet-to-be-regulated behaviors like Level 2 systems off-highway, system-initiated lane changes with hands-off the wheel etc.  Some of these regulations are outdated and rules-based, which makes FSD illegal in its current form. Changing FSD to be compliant with these rules would make it unsafe and unusable in many cases. While we have changed FSD to be maximally compliant where it is logical and reasonable, we won't sacrifice the safety of a proven system or materially deteriorate customer usability.  As a result, we are gathering evidence to get exemptions on a specific rule-by-rule basis. Unfortunately, the real world fleet-proven safety wins alone are considered insufficient.  Currently, RDW has committed to granting Netherlands National approval in February 2026. Please contact them via link below to express your excitement & thank them for making this happen as soon as possible. Upon NL National approval, other EU countries can immediately recognize the exemption and also allow rollout within their country. Then we will bring it to a TCMV vote for official EU-wide approval.  We're excited to bring FSD to our owners in Europe soon! rdw.nl/en/contact/our-contac…
806
1,801
572
10,248
4,148,715