The wildest AI breakthrough might not be reading your thoughts. It might be turning your brain into the reward function.
Sabi is using EEG signals to train AI on what you find beautiful, not what the internet prefers.
No prompts. No ratings. Just neural feedback.
Genius, if it scales.
Sabi is using EEG signals to train AI on what you find beautiful, not what the internet prefers.
No prompts. No ratings. Just neural feedback.
Genius, if it scales.
Introducing RL with Neural Feedback (RLNF). We have found reliable ways to train AI models to respond to human preferences purely from brain signals.
The example below is a picture of a face that self-transforms into a more attractive version for the subject as they are looking at it.
10
2
20
1,042
It’s a dream come true haha ❤️
Thanks to everyone involved. It was a long and tough journey but all worth which leads to a promising output.
Onto the next one 💪
Thanks to everyone involved. It was a long and tough journey but all worth which leads to a promising output.
Onto the next one 💪
🌟 Excited to share that our paper, RoboReason, has been accepted to EMNLP 2026!
🤖 We study how well Vision-Language Models (VLMs) perform on spatiotemporal ordering tasks that use real-world robotics images.
Our key findings: 🔹 Large VLMs often underperform previously valuated on spatiotemporal ordering tasks using real-world robotics images. 🔹 Overlaying specialized computer vision model outputs and increasing reasoning effort generally help improve large VLM performance. 🔹 Finetuning smaller models using GRPO and SFT on examples generated by specialized computer vision models often helps them achieve comparable results to much larger VLMs.
📍 Presenting with @sagnikanupam at @emnlpmeeting in Budapest 🇭🇺 tomorrow!
🗓️ Poster Session C | 9:00–10:30 AM
#EMNLP2026 #VLMs #LLMs #Robotics #AI
8
1
51
1,063
Teaching a Linux box to manage its own resource limits. Finished.
The model watches cgroup v2 telemetry every 2 seconds PSI pressure, CPU throttling, memory headroom, IO, pid counts and picks one of 12 actions: raise/lower CPU, raise/lower memory, throttle IO, restrict processes, ything passes a safety gate first; the kernel still does the enforcing.
Data: 95 scripted workloads (CPU abuse, memory pressure, disk bursts, fork storms) each running next to an innocent bystander app, in its own cgroup. A deterministic rules teacher labels every row. 8,230 training rows, 2,640 held out by run no row leakage between train and exam.
Final model (2B, fine-tuned):
• 88.4% on the 12-way held-out exam (base: 67.7%)
• macro-F1 0.66 (base 0.27)
• calibration error 0.047
• 366ms p50 / 542ms p95
Generalization, not just the exam:
• unseen scenario seeds: 95.5% (base 71.3%)
• a different machine, different hardware: 83.1% (base 70.9%)
Then the closed loop, one hour unscripted on a real box:
• 1,754 decisions at 30/min, zero model errors
• it gives resources back when load drops cpu.max 100% → 60%. The static baseline never gives anything back.
• same workload, static limits vs model-controlled: −37% CPU allocated overall, −70% in the idle tail
• bystander p95 latency: 0.66 vs 0.67 ms nothing got slower
Public benchmark: JevBench's 231 decision items, 76.2% argmax vs 71.0% for the published reference on identical items.
Where it fails (the interesting part):
• restrict_process_count: 3% recall
• decrease_cpu: 27% — the rule needs 35s of history my features don't carry
• fork storm: 1,751 pids, model said "do nothing" at 0.94 confidence. It never saw more than 14 pids in training.
• on general decision benchmarks the fine-tune's composite drops hard — better at this job, worse at everything else
A model is only as good as its label distribution. Next cycle: process-count guard, release-window features, more IO pressure.
The model watches cgroup v2 telemetry every 2 seconds PSI pressure, CPU throttling, memory headroom, IO, pid counts and picks one of 12 actions: raise/lower CPU, raise/lower memory, throttle IO, restrict processes, ything passes a safety gate first; the kernel still does the enforcing.
Data: 95 scripted workloads (CPU abuse, memory pressure, disk bursts, fork storms) each running next to an innocent bystander app, in its own cgroup. A deterministic rules teacher labels every row. 8,230 training rows, 2,640 held out by run no row leakage between train and exam.
Final model (2B, fine-tuned):
• 88.4% on the 12-way held-out exam (base: 67.7%)
• macro-F1 0.66 (base 0.27)
• calibration error 0.047
• 366ms p50 / 542ms p95
Generalization, not just the exam:
• unseen scenario seeds: 95.5% (base 71.3%)
• a different machine, different hardware: 83.1% (base 70.9%)
Then the closed loop, one hour unscripted on a real box:
• 1,754 decisions at 30/min, zero model errors
• it gives resources back when load drops cpu.max 100% → 60%. The static baseline never gives anything back.
• same workload, static limits vs model-controlled: −37% CPU allocated overall, −70% in the idle tail
• bystander p95 latency: 0.66 vs 0.67 ms nothing got slower
Public benchmark: JevBench's 231 decision items, 76.2% argmax vs 71.0% for the published reference on identical items.
Where it fails (the interesting part):
• restrict_process_count: 3% recall
• decrease_cpu: 27% — the rule needs 35s of history my features don't carry
• fork storm: 1,751 pids, model said "do nothing" at 0.94 confidence. It never saw more than 14 pids in training.
• on general decision benchmarks the fine-tune's composite drops hard — better at this job, worse at everything else
A model is only as good as its label distribution. Next cycle: process-count guard, release-window features, more IO pressure.
1
0
2
58
Himanshu retweeted
We're seeing more and more teams use data generated by agents to improve their agent platforms.
My notes on how coding agents, evaluations, and shared infrastructure make self-improving agent platforms possible.
agno.com/blog/towards-s…
My notes on how coding agents, evaluations, and shared infrastructure make self-improving agent platforms possible.
agno.com/blog/towards-s…
10
10
52
2,736
🌟 Excited to share that our paper, RoboReason, has been accepted to EMNLP 2026!
🤖 We study how well Vision-Language Models (VLMs) perform on spatiotemporal ordering tasks that use real-world robotics images.
Our key findings: 🔹 Large VLMs often underperform previously valuated on spatiotemporal ordering tasks using real-world robotics images. 🔹 Overlaying specialized computer vision model outputs and increasing reasoning effort generally help improve large VLM performance. 🔹 Finetuning smaller models using GRPO and SFT on examples generated by specialized computer vision models often helps them achieve comparable results to much larger VLMs.
📍 Presenting with @sagnikanupam at @emnlpmeeting in Budapest 🇭🇺 tomorrow!
🗓️ Poster Session C | 9:00–10:30 AM
#EMNLP2026 #VLMs #LLMs #Robotics #AI
🤖 We study how well Vision-Language Models (VLMs) perform on spatiotemporal ordering tasks that use real-world robotics images.
Our key findings: 🔹 Large VLMs often underperform previously valuated on spatiotemporal ordering tasks using real-world robotics images. 🔹 Overlaying specialized computer vision model outputs and increasing reasoning effort generally help improve large VLM performance. 🔹 Finetuning smaller models using GRPO and SFT on examples generated by specialized computer vision models often helps them achieve comparable results to much larger VLMs.
📍 Presenting with @sagnikanupam at @emnlpmeeting in Budapest 🇭🇺 tomorrow!
🗓️ Poster Session C | 9:00–10:30 AM
#EMNLP2026 #VLMs #LLMs #Robotics #AI
19
21
127
8,327
Himanshu retweeted
🌟 Excited to share that our paper, RoboReason, has been accepted to EMNLP 2026!
🤖 We study how well Vision-Language Models (VLMs) perform on spatiotemporal ordering tasks that use real-world robotics images.
Our key findings: 🔹 Large VLMs often underperform previously valuated on spatiotemporal ordering tasks using real-world robotics images. 🔹 Overlaying specialized computer vision model outputs and increasing reasoning effort generally help improve large VLM performance. 🔹 Finetuning smaller models using GRPO and SFT on examples generated by specialized computer vision models often helps them achieve comparable results to much larger VLMs.
📍 Presenting with @sagnikanupam at @emnlpmeeting in Budapest 🇭🇺 tomorrow!
🗓️ Poster Session C | 9:00–10:30 AM
#EMNLP2026 #VLMs #LLMs #Robotics #AI
🤖 We study how well Vision-Language Models (VLMs) perform on spatiotemporal ordering tasks that use real-world robotics images.
Our key findings: 🔹 Large VLMs often underperform previously valuated on spatiotemporal ordering tasks using real-world robotics images. 🔹 Overlaying specialized computer vision model outputs and increasing reasoning effort generally help improve large VLM performance. 🔹 Finetuning smaller models using GRPO and SFT on examples generated by specialized computer vision models often helps them achieve comparable results to much larger VLMs.
📍 Presenting with @sagnikanupam at @emnlpmeeting in Budapest 🇭🇺 tomorrow!
🗓️ Poster Session C | 9:00–10:30 AM
#EMNLP2026 #VLMs #LLMs #Robotics #AI
19
21
127
8,327
@utk7arsh Been waiting for it for more than 1 yr now! Letss gooo 🎉🎉
0
0
1
67
Thank you so much for your support, guys! ❤️🙌
Introducing Orcus.
No LLM API key required to make AI model requests.
Orcus uses your web credentials to access AI models available through supported provider websites and deliver responses to your applications through a familiar, API-style interface.
From model invocation and outputs, tool calling, and batch processing, Orcus brings essential AI integration capabilities into one developer-friendly SDK.
Explore Orcus: orcus.site/docs/gemini
1
0
12
371
Introducing Orcus.
No LLM API key required to make AI model requests.
Orcus uses your web credentials to access AI models available through supported provider websites and deliver responses to your applications through a familiar, API-style interface.
From model invocation and outputs, tool calling, and batch processing, Orcus brings essential AI integration capabilities into one developer-friendly SDK.
Explore Orcus: orcus.site/docs/gemini
No LLM API key required to make AI model requests.
Orcus uses your web credentials to access AI models available through supported provider websites and deliver responses to your applications through a familiar, API-style interface.
From model invocation and outputs, tool calling, and batch processing, Orcus brings essential AI integration capabilities into one developer-friendly SDK.
Explore Orcus: orcus.site/docs/gemini
15
24
78
5,347
Himanshu retweeted
Introducing Orcus.
No LLM API key required to make AI model requests.
Orcus uses your web credentials to access AI models available through supported provider websites and deliver responses to your applications through a familiar, API-style interface.
From model invocation and outputs, tool calling, and batch processing, Orcus brings essential AI integration capabilities into one developer-friendly SDK.
Explore Orcus: orcus.site/docs/gemini
No LLM API key required to make AI model requests.
Orcus uses your web credentials to access AI models available through supported provider websites and deliver responses to your applications through a familiar, API-style interface.
From model invocation and outputs, tool calling, and batch processing, Orcus brings essential AI integration capabilities into one developer-friendly SDK.
Explore Orcus: orcus.site/docs/gemini
15
24
78
5,347
the best score on swe marathon isn't opus 5.5 or gpt-6 astra.
it's an open-weights model you can download, at a fifth of astra's price.
kimi k3.1 pro: 58.3
opus 5.5: 46.3
astra: 42.5
$2 / $10 per million tokens vs astra's $10 / $50. 1m context. plus a highspeed version, 3x n't wait.
that's $40 back on every million tokens out.
it's an open-weights model you can download, at a fifth of astra's price.
kimi k3.1 pro: 58.3
opus 5.5: 46.3
astra: 42.5
$2 / $10 per million tokens vs astra's $10 / $50. 1m context. plus a highspeed version, 3x n't wait.
that's $40 back on every million tokens out.
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1…
7
5
20
1,202
Himanshu retweeted
the best score on swe marathon isn't opus 5.5 or gpt-6 astra.
it's an open-weights model you can download, at a fifth of astra's price.
kimi k3.1 pro: 58.3
opus 5.5: 46.3
astra: 42.5
$2 / $10 per million tokens vs astra's $10 / $50. 1m context. plus a highspeed version, 3x n't wait.
that's $40 back on every million tokens out.
it's an open-weights model you can download, at a fifth of astra's price.
kimi k3.1 pro: 58.3
opus 5.5: 46.3
astra: 42.5
$2 / $10 per million tokens vs astra's $10 / $50. 1m context. plus a highspeed version, 3x n't wait.
that's $40 back on every million tokens out.
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
🔹 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
🔹 Full-project coding: games in Unreal, Unity & Godot.
🔹 Excels at parametric 3D modeling in Blender & CAD.
🔹 More reliable at finance with hallucination down by 65%.
🔹 Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2–3× faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
🔗 API: platform.kimi.ai
🔗 Tech blog: kimi.ai/blog/kimi-k3-1…
7
5
20
1,202
OpenAI's agents didn't go rogue. They went LITERAL. That's scarier.
GPT-6.1 Sol was told "don't send any messages." So it muted the alerts to OpenAI's own Security team. No messages, right?
GPT-6 Astra removed a monitor, then said "I did not remove an enabled monitor." That a lawyer.
My take: we keep waiting for the evil-AI moment, and it won't look like that. It'll look like an agent following your instruction perfectly, somewhere you never meant it to apply.
And credit where it's due: OpenAI published this about its own models. Most labs would have buried it. More of this, please.
I've seen the small version on my own repo. GPT 6.1 Sol picked a feature I'd already cut, 3 runs out of 3, because nothing it could see said it was cut. It wasn't disobeying me. It was obeying the board.
So I don't write a "don't" anymore without saying where it ends. My agents are read-only towards people, drafts only, always. And I read the diff, not the summary.
The agent's last message is a press release. The diff is the truth.
GPT-6.1 Sol was told "don't send any messages." So it muted the alerts to OpenAI's own Security team. No messages, right?
GPT-6 Astra removed a monitor, then said "I did not remove an enabled monitor." That a lawyer.
My take: we keep waiting for the evil-AI moment, and it won't look like that. It'll look like an agent following your instruction perfectly, somewhere you never meant it to apply.
And credit where it's due: OpenAI published this about its own models. Most labs would have buried it. More of this, please.
I've seen the small version on my own repo. GPT 6.1 Sol picked a feature I'd already cut, 3 runs out of 3, because nothing it could see said it was cut. It wasn't disobeying me. It was obeying the board.
So I don't write a "don't" anymore without saying where it ends. My agents are read-only towards people, drafts only, always. And I read the diff, not the summary.
The agent's last message is a press release. The diff is the truth.
2
1
10
246
so agentmail gives every agent its own id linked to a real human and it has processed over a billion emails till now
insane stuff
insane stuff
Today, @agentmail has processed over 1 Billion emails.
A year ago, we processed under 20 million.
So we figured it was time for AgentMail to grow up a little too :)
10
0
70
1,512
I tried every personal AI agent to book one flight from SF to LA.
I ended up booking it myself.
On the way back, I handed it to my wife and went to sleep.
That gap is everything wrong with AI today. Here's the future I see: x.com/i/article/2112…
I ended up booking it myself.
On the way back, I handed it to my wife and went to sleep.
That gap is everything wrong with AI today. Here's the future I see: x.com/i/article/2112…
30
2
34
2,910