steer L6 N7448 -0.8, L5 N6203 -0.75 and L25 N8324 -0.4 on qwen3.5 4b
now you have golden gate qwen
same recipe as evil qwen by @_lyraaaa_
now you have golden gate qwen
same recipe as evil qwen by @_lyraaaa_
Can someone please make a Golden Gate Qwen
5
2
83
so theo has a repo called fleet to sync his skills and tools across machines, and i built a tool called fleet to ALSO do something similar + running stuff across machines, giving agents access to proper tools like computer use across a fleet of machines (using cua), everything is we're all using the same tools and we're all hitting the same weird agentic friction where the models are insanely capable of doing things across insane time horizons and dozens of devices at once for a task, but just no tools to do it right/quickly/easily depending on the harness they're in
this repo is a personal lovechild of mine and has ridiculous stuff like webcam access for your machines and ANDROID COMPUTER USE because i wanted to iterate faster on building my own android apps, and probably some version of iOS Computer Use soon
ofc, use at your own risk, ive done a thorough security review, but file issues and PRs, this is pretty much allowing code execution across your devices and that has it's risks, but it should be pretty much the same as giving your agents ssh access to your machines, just all in one place with a skill attached to guide them through doing it simply
the goal is to be harness agnostic, so it's one CLI for all called using your harness's bash tool (mcp available, but not recommended because of just tool bloat), t3code does most of what my fleet cli does BY DEFAULT, this is a cli tool IF you don't use t3. you can use these in parallel with ruddr (another agent orchestration cli I made) for much better cross machine orchestration
repo below!
this repo is a personal lovechild of mine and has ridiculous stuff like webcam access for your machines and ANDROID COMPUTER USE because i wanted to iterate faster on building my own android apps, and probably some version of iOS Computer Use soon
ofc, use at your own risk, ive done a thorough security review, but file issues and PRs, this is pretty much allowing code execution across your devices and that has it's risks, but it should be pretty much the same as giving your agents ssh access to your machines, just all in one place with a skill attached to guide them through doing it simply
the goal is to be harness agnostic, so it's one CLI for all called using your harness's bash tool (mcp available, but not recommended because of just tool bloat), t3code does most of what my fleet cli does BY DEFAULT, this is a cli tool IF you don't use t3. you can use these in parallel with ruddr (another agent orchestration cli I made) for much better cross machine orchestration
repo below!
13
0
100
my life after getting access to abliterated deepseek
what the fuck why wasn't i just doing this the whole time
glm 5.3 + deepseek v4.1 flash is an insane cybersec combo
thanks @bedros_p for putting me on this
10
17
829
lmao i made more progress with $2.5 and abliterated deepseek (thanks @skeptrune and @morphllm for the access) in 2 hours than i did with 2 days of frontier models because of their shitty cyber guardrails
goodnight
goodnight
11
5
205
the fuck do you mean you WON'T train the model properly and entirely just drop it? why? seems very wasteful
if a model shows any misaligned behaviour you just kill the run and go back to a previous checkpoint?
hope they know what they're doing
if a model shows any misaligned behaviour you just kill the run and go back to a previous checkpoint?
hope they know what they're doing
Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.
On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.
While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.
After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.
openai.com/index/securing…
3
1
31
Kimi K3.1 Pro BEATS GPT-6 Astra and Claude Fable 5.1 on the Artificial Analysis Intelligence Index to take 4th place on the leaderboard!
5
4
57
Mistral Large 4 collection was created on HuggingFace 4 minutes ago, the models weights release is imminent.
2
1
38
what the fuck why wasn't i just doing this the whole time
glm 5.3 + deepseek v4.1 flash is an insane cybersec combo
thanks @bedros_p for putting me on this
glm 5.3 + deepseek v4.1 flash is an insane cybersec combo
thanks @bedros_p for putting me on this
64
70
1,918
im losing my mind with all these refusals
i do not want to pay big boy money for @abliteration_ai
other glm 5.3 (flash) abliterated providers are weird
i can only run 120b param models and im not finding any new vulnerabilities and am barely confident in my patches
what do?
i do not want to pay big boy money for @abliteration_ai
other glm 5.3 (flash) abliterated providers are weird
i can only run 120b param models and im not finding any new vulnerabilities and am barely confident in my patches
what do?
67
7
401
@shaurizard i might need a little more than 30 minutes
profiling and benching is a bitch but im going to burn thru subagents and let them lock some slots for experiments because i dont have any more ideas other than these last 3-4
and hey i have a pretty sick decomp toolset and skills now
profiling and benching is a bitch but im going to burn thru subagents and let them lock some slots for experiments because i dont have any more ideas other than these last 3-4
and hey i have a pretty sick decomp toolset and skills now
1
0
1
holy frontier AI lab mog, full speed ahead with their IPO huh?
Anthropic is THE FRONTIER AI LAB, safe to say they won the Stage 1 of the AI race?
Fable 5.5 has some extremely big shoes to fill.
Anthropic is THE FRONTIER AI LAB, safe to say they won the Stage 1 of the AI race?
Fable 5.5 has some extremely big shoes to fill.
Anthropic is a Public Benefit Corporation. Today, we’re publishing our annual Public Benefit Report. This report shares our five pillars and takes stock of the actions we’re taking to advance them.
Read the full report at this page: anthropic.com/public-benefit…
5
0
105
claude stickers!! while it recompiles project gotham racing 4 to dx12 on pc in the background
5
0
41
my CTO asked 'Council of Astra' (aka Astra and dozens of Astra subagents) to come up with a moonshot of an idea for our startup and it told us to pivot to becoming an Airbnb clone and burnt 40% of his weekly quota
(we are a tech/AI-native company)
what is openai doing bro
(we are a tech/AI-native company)
what is openai doing bro
1
0
16
I have tokens to burn because my Claude reset's gonna expire, what do I mess with/throw tokens at to solve?
5
0
21
Lisan and Teor are the same person, Alexandr Wang.
@teortaxesTex Oh yeah forgot. Here is it again
1
0
21
@scaling01 cmon man wdym it's less efficient
tokens don't matter where we're going
x.com/ValsAI/status/…
tokens don't matter where we're going
x.com/ValsAI/status/…
The implication here is that the offensive capability previously limited to a closed-source, trusted access model is now available to anyone via the latest DeepSeek model. In practice, DeepSeek V4.1 Flash 1017 spends significantly more tokens but is considerably cheaper per-token so achieves the same outcome at a far lower total price.
1
0
13
DeepSeek seems to have dropped support for their Hopper stack entirely (H100/H200) and only support Blackwell and Huawei Ascend NPUs with their new open sourced inference framework.
Big China victory.
Big China victory.
🚀 Introducing DeepInfer, our in-house inference framework, now fully open source!
Powering DeepSeek-V4.1-Flash at scale with high throughput and low latency. We’re sharing our production stack to help the community lower compute costs, reduce energy use, and make AI more affordable.
🔗 github.com/deepseek-ai/De…
1/4
8
9
198
MiniMax M3.1 is moving out of preview soon and will be generally available, weights are likely coming today.
2
3
82
bro thinks cloudflare turnstile would've prevented the security breach
why do we have so many larpers in tech
why do we have so many larpers in tech
7
1
149
DeepSeek has built their own search engine with a lot of filtering to avoid modern SEO/GEO spam, and powers their new search and fetch APIs with it!
Each API costs $0.75 per 1,000 successful requests, free if you use DeepSeek in Claude Code, Codex or DeepSeek harness until Nov 1
Each API costs $0.75 per 1,000 successful requests, free if you use DeepSeek in Claude Code, Codex or DeepSeek harness until Nov 1
52
75
1,594
i lost my rebase streak
my life is meaningless
you can see my google of shame as i try to remember where the fuck the caesar salad originated from
goodnight im going to be so hungover tomorrow
my life is meaningless
you can see my google of shame as i try to remember where the fuck the caesar salad originated from
goodnight im going to be so hungover tomorrow
8
0
28
Muse Spark Open will be available on Groq, Together, Fireworks, Baseten and Nebius Token Factory soon!
An actually somewhat modern, omnimodal model running on super fast Groq is something I'm very excited for!
An actually somewhat modern, omnimodal model running on super fast Groq is something I'm very excited for!
1
0
13
NVIDIA Nemotron 3.5 Bolt 74B is launching soon, has 9 billion active params and will fit on a single 3090 with a 2-bit QAT variant!
0
2
28
rebase is back!
live trivia on your phone, this friday 3pm est
get 5 friends to sign up in 10 minutes and get $10
download the app now → rebase.tv
0
0
10
it's also... meh in terms of code quality (judged by Opus and Sol below)
only benefit is that it's changes are usually extremely small, but moving into the world of AI recomps and decomps this is too little too late unfortunately
Argon is benchmaxxed, expensive and subpar.
only benefit is that it's changes are usually extremely small, but moving into the world of AI recomps and decomps this is too little too late unfortunately
Argon is benchmaxxed, expensive and subpar.
Gemini 4 Argon is extremely inefficient and slow in my early testing.
Does it get the task done? Yes.
But at over 3x the cost while already costing half the other models during the discounted period, it doesn't seem like a super efficient release :/
9
5
128
Gemini 4 Argon is extremely inefficient and slow in my early testing.
Does it get the task done? Yes.
But at over 3x the cost while already costing half the other models during the discounted period, it doesn't seem like a super efficient release :/
Does it get the task done? Yes.
But at over 3x the cost while already costing half the other models during the discounted period, it doesn't seem like a super efficient release :/
13
3
270
