Offensive Security Certified Professional (OSCP) passionate about option trading, cybersecurity, AI and new technologies.
Canada
Joined February 2009
- Tweets745
- Following450
- Followers160
- Likes3.7K
This is not perfect. Yet even an imperfect, AI movie of Arthas can give me actual shivers. Insane that Blizzard never made an actual movie out of Arthas' story. youtu.be/PYKDyXnJl5E
Jeff Noël | OSCP retweeted
What 🫣🫣
try.cloudflare.com/
What a time to be alive man….
Why is there no benchmarks testing OpenAI's Pro thinking level versus Max/Extra High?
I assume the cost would be through the roof, but I'd be curious to see the intelligence difference.
@thsottiaux
Jeff Noël | OSCP retweeted
CancerBench: the frontier model cancer cure benchmark.
AI lab CEOs keep talking about curing cancer, so I made a benchmark.
One metric: how many types of cancer has your model cured?
All models are currently tied at zero.
It’s time to hillclimb!
cancerbench.com
The amount of emdash in X comments is unsettling. Are you guys all farming with bots all day long?
What about interactions?
The next big step IMO is getting Fable 5.1 or even GPT-6 Astra level of cognition and intelligence with actual usage.
The models are great, but the current 20x subscriptions weekly limits on both get completely used in a SINGLE day.
Make it make sense.
Fable 5.1 is insane, but it's eating through usage limits like crazy. Roughly ~1 hour in a 5 hours session post-reset, nothing fancy 3-5 subagents in a regular workflow, on Fable 5.1 XHigh, then lowered to Medium with Opus (ugh) and Sonnet subagents:
I wonder if @AnthropicAI told their employees to go silent on X and ignore everyone. Clearly they don't care about the feedback on here since Opus 5 release and the recent outages.
While everyone's asking for whatever, all I want to know is, were the recent breakthroughs and improvements thanks to 5.7, 6.0, or a new mythos-class model?
@thsottiaux
Anthropic better drop some goodies soon or many $200 subs will be moving away from them. While they delivered a terrible AX (agentic experience) model through Opus 5, @OpenAI @OpenAIDevs has been improving very quickly and they're actively interacting with the community as well.
Jeff Noël | OSCP retweeted
End of an era… 💔 After 14 years and 1.7 billion views, YouTube removed the original Rick Roll. 😢
youtu.be/dQw4w9WgXcQ?si=c9DW…
Let your friends know.
People should code more tools oriented towards agents.
Currently, most agents are using tools designed for humans as well (grep, rg, etc.)
Anyone has been vibecoding/researching stuff meant and optimized for AI explicitely?
It is deliberately conservative:
• dry-run by default
• local transcript parsing
• secret redaction
• compact digests instead of raw 40MB transcripts
• no source-code edits
• no autonomous skill rewrites
• restricted headless permissions
I also added a pre-registered 14-day kill test.
Re-derivation must decrease.
Rule precision must stay ≥80%.
Rules don't make sense if they're useless.
The time saved must justify the cost.
If those conditions fail, the experiment stops.
No moving the goalposts.
nightly-consolidation is open source and MIT licensed.
No fine-tuning. No vector database. No separate memory service.
Just Claude Code, Python, local transcripts and durable evidence-backed rules.
github.com/JF10R/nightly-con…
Claude Code can repeat the same mistake for months.
I analyzed 800+ real sessions. 73% hit an error signature already seen before.
The agent was working every day, but its experience was not compounding.
So I built nightly-consolidation. 🧵
