ノーネーム retweeted
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…
Having enough money to buy anything and live without ever needing to work again while that money just keeps growing feels both good and wrong / unfair at the same time.
That unfairness feeling kicks in only when you fully stop working because you can no longer tell yourself that you deserve it because of your hard work.
And you can’t tell about this to almost anyone because they would view it as humble bragging.
ノーネーム retweeted
Key takeaways:
1. Screenshots beat prompts. One image holds fonts, colors, layout, icons. Gemini 3 captures it instantly. Stop writing 1000-word descriptions.
2. Hero section = 50% of your time. Use Superhero to collect hero references. It's your cover image for social—everything else is secondary.
3. Template business prints money. Sell on Webflow, Framer, UI8 for $50-100 each. Stack enough and you're clearing six figures.
4. Taste is the new moat. AI raises the baseline. Your visual judgment becomes the differentiator, not your Figma speed.
5. Avoid AI slop. Purple gradients, Lucid icons, generic layouts. Switch to Iconify Solar for variety—it has outline, broken, duotone styles.
6. Use Simple Icons via Iconify for logos. Apple, Google, whatever. One API reference in your prompt, no more hunting SVGs.
7. Work section by section. Hero first in Aura/v0/Cursor, then features, then footer. Faster iterations, better creative control.
8. Remix everything. Grab screenshots from Mobbin (sites → sections) or Bento Grids. Change colors, text, animations. Mold it new.
9. Use negative prompts. "Don't change anything else" or "keep the hero" prevents AI from breaking what already works.
10. Fix images manually. AI breaks them constantly. Use Midjourney or Nano Banana Pro to generate replacements and mockups.
11. Typography matters. Google Fonts are free and AI-friendly. Inter is overused—try Newsreader or Playfair Display for differentiation.
12. Add craft signals. 01/02/03 numbering, beam animations, noodles, container-sized grid lines. Use Unicorn Studio for hero backgrounds.
13. Steal headlines from H1 Gallery. Better hooks = better conversions. cta.gallery for call-to-action inspiration.
14. Present work properly. Use Screen Studio for recordings. Screenshot with polished backgrounds—this is what drives social engagement.
15. Speed trick: Use GPT-5.1 for small text/color changes. Save Gemini 3 for complex UI generation.
Bottom line: Superhero for heroes, Mobbin for sections, Bento Grids for cards, Iconify for icons. Screenshot → Gemini 3 → Remix → Polish. Your taste curates. AI executes.
New LLM benchmark just dropped
Exploit-bench, aka SCONE-bench red.anthropic.com/2025/smart…
ノーネーム retweeted
The EU Council appears ready to approve Chat Control.
They're trying to pass an even more far-reaching Chat Control law through the back door, in a form even more intrusive than the originally rejected plan, without needing any EU countries' votes.
ノーネーム retweeted
The EU Council appear ready to approve Chat Control. This must be stopped. To highlight the corruption behind the proposal, Mullvad VPN now present "And Then?"
The backstory: mullvad.net/blog/mullvad-vpn…
The dude keeps on scheming. And people keep on buying. Never learning. Always eventually getting rugged.
EXCLUSIVE: Andre Cronje's Flying Tulip raises $200 million at $1 billion token valuation theblock.co/post/372787/andr…
ノーネーム retweeted
“OpenAI signed a $300 Billion GPU data center deal with Oracle.”
> Larry Ellison gained $101B in one day
> now the richest man alive.
> but Sam doesn't have $300B
> Larry doesn't have $300B worth of GPUs.
We’re living in a simulation.
Will social media wealth flaunting lose all its value with all the AI image/video tools getting indistinguishable from reality?
Powerful mental model:
Seek out rejection. Push for it. Fast.
If you don’t get rejected, the resulting outcome is just a side effect.
Goal is to get rejected.
Eliminates all fear of rejection and actually makes you less likely to fail.
ノーネーム retweeted
New Google Play Store policy forces AML/KYC on non-custodial wallets in the US, effectively bans non-custodial wallet developers from Play Store in EU
Full Story👇
therage.co/google-play-store…
Who buys this dogshit
Figma, after its first day of trading, became the most expensive U.S. tech company by Forward Price/Earnings valuation multiple (602x). IPO mania is back. $FIG
