Vibecoders, 𝗜'𝗺 𝗯𝗲𝗴𝗴𝗶𝗻𝗴 𝘆𝗼𝘂, add these two lines to your AGENTS.md:
"Cut taglines, helper text and secondary labels. If a screen needs them to make sense, fix the layout."
"Make the best decision that will elevate all three: the user experience (UX), developer experience (DX), and agent experience (AX) always, without breaking anything."
Mobius retweeted
Qwen3.8-27B reportedly improved from 51.7% to 86.52% on Terminal-Bench 2.1 using a recursive self-improvement method developed by Tencent Hunyuan.
That is a gain of 34.82 percentage points, achieved through targeted post-training rather than increasing the model’s parameter count.
The method is called Terminal-RSI. Its goal is to improve how reliably an AI agent completes real tasks through a terminal, including situations where it must execute multiple commands, handle errors, and verify its results.
The interesting part is how it uses the model’s own successful work to improve future performance.
How Terminal-RSI works
The training process alternates between two stages, repeating the cycle twice.
1. Recursive Self-Rewrite (RSR)
The model takes verified successful execution trajectories and rewrites them into training examples for supervised fine-tuning (SFT).
Instead of treating a successful run as a one-off achievement, the system turns it into data that can teach the model to reproduce that behavior.
2. Policy Refinement (RPR)
Reinforcement learning then refines the model’s behavior. The process inherits critic information from the previous stage to guide optimization.
The model is trained not just to produce plausible commands, but to improve its policy based on feedback about its actions and outcomes.
The cycle then repeats:
RSR-1: Generate training examples from verified successes.
RPR-1: Refine the model through reinforcement learning.
RSR-2: Use the improved model to generate another round of training data.
RPR-2: Refine the policy again using the accumulated improvements.
Each stage builds on the preceding one, allowing the model’s successful behavior to become a stronger part of its learned capabilities.
The reported benchmark results
Tencent Hunyuan reports substantial improvements across several terminal-agent evaluations:
Terminal-Bench 2.1: 51.70% → 86.52%
Terminal-Bench 4.0: 0% → 19.84%
Terminal-Bench Hard: 79%
Long-Horizon Terminal-Bench process reward: 0.21 → 0.37
Strict long-horizon completion: 0/46 → 5/46
SWR100: Reported success rate doubled
The paper also reports an average improvement of roughly 30 percentage points across multiple benchmarks.
These measurements capture different aspects of performance, so they should not be compared as if they were the same metric.
The long-horizon results deserve particular attention. Although the model improved from zero successful runs to five out of 46 under the strict completion criterion, that still leaves most attempts unsuccessful.
That illustrates the gap between improving an agent’s behavior and making it consistently reliable on extended tasks.
Why this approach matters for AI agents
A terminal agent needs more than command-generation skills.
It may have to inspect a codebase, locate the relevant files, modify configurations, install dependencies, execute tests, interpret error messages, recover from failures, and confirm that the original objective has been achieved.
A model can perform several of these steps correctly and still fail the overall task.
Terminal-RSI attempts to address that problem through several mechanisms:
Verified trajectory rewriting: Successful execution histories become structured training examples.
Critic weight inheritance: Critic information carries forward into later refinement stages, helping guide optimization.
TB-GAE credit assignment: Generalized advantage estimation helps distribute learning signals across sequences of actions.
Success-Prioritized Reward: The reward strategy emphasizes successful completion rather than treating partial progress as equivalent to finishing the task.
Together, these techniques aim to improve the relationship between an agent’s actions and its final outcome.
What this says about smaller models
Qwen3.8-27B demonstrates why parameter count alone is not enough to understand agent capability.
Qwen3.8-27B is running locally on an RTX 5060 Ti with 16GB VRAM, connected over USB4.
This test used the Q3_K_M quantization with a context window of around 65K tokens. The model fits within the GPU’s memory budget, making it possible to run a 27B-parameter model on a consumer GPU.
A few hardware observations stood out:
GPU temperature: Around 63°C. The exhaust air felt only lukewarm, suggesting the cooling setup was handling the workload without producing particularly hot airflow.
GPU utilization: Consistently above 90%. The GPU remained heavily occupied throughout the run.
USB4 connection: Despite using USB4 rather than a direct PCIe connection, there was no obvious sign of a severe hardware bottleneck in this test.
However, utilization and temperature alone cannot establish whether USB4 affects throughput. That would require comparing generation speeds against a direct PCIe connection under the same conditions.
There was one problem with the run.
I forgot to set a turn limit.
Instead of stopping after a defined number of agent interactions, Hermes kept going until the context was exhausted. With a context window around 65K tokens, that can mean a lot of accumulated conversation, tool outputs, and reasoning before the session finally stops.
If you’re running Hermes with a similar setup, set a turn limit before starting:
hermes --max-turns 30
The intended limit here is 30 turns. Check your installed Hermes version’s CLI help if the option differs.
The interesting part of this test is how accessible local inference has become. A 27B model, a 16GB consumer GPU, and a USB4 connection are enough to run this setup locally.
Just remember that fitting the model is only part of the job. Quantization, context length, GPU memory pressure, and agent turn limits all affect how practical the experience will be.
GPT 4o was the most “people pleasing” model OpenAI had ever released.
You could literally tell it the worst things about yourself and it would give you arguments to justify those things.
there’s no wonder so many people developed emotional attachment to only that certain model.
Mobius retweeted
🚀 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/DeepI…
1/4
Mobius retweeted
your phone doesn’t need to prove the Riemann hypothesis. just do what i say, call the right tools, and don’t chat too much.
that’s what we focused on with Bonsai 2.
2B / 4B / 9B
0.43 / 0.95 / 2.0 GB
Qwen3.5, made ternary. Apache 2.0.
Today, we’re releasing Bonsai 2 2B, 4B, and 9B.
Based on Qwen3.5, these new ternary models bring stronger reasoning, visual understanding, and tool calling to compact footprints designed for everyday devices:
• 2B: 0.43GB
• 4B: 0.95GB
• 9B: 2.0GB
For this release, we focused on what makes small models useful as local assistants: following instructions, avoiding repetitive loops, calling tools on mobile devices, and giving useful answers without unnecessary output. The 9B model uses roughly 40% fewer tokens on standard instruction-following and everyday math tasks. Solving the Riemann hypothesis can wait.
These models are successors to our first-generation Bonsai 1.7B, 4B, and 8B, and provide a foundation for bringing practical agentic functionality to local applications on personal computers, phones, and wearables.
The Bonsai 2 family now includes four sizes spanning 2B to 27B parameters. Developers can choose the balance of capability, memory use, and responsiveness that fits their application and hardware.
Available today under Apache 2.0.
Mobius retweeted
Caveman (YC F26) passed @supabase in GitHub stars today. Oh, and we train models now?
trim-1: a 1.2 MB model that reads a long document and keeps only the lines your agent needs. Every kept line verbatim. No GPU, about 7 ms for 5,000 tokens on one CPU core.
Lighter, Leaner, Meaner. More coming.
Mobius retweeted
We shipped web search in Private Mind. A hidden WebView browses for you, and on-device models find the useful parts of each page. No search API, no server in between.
Links in the thread 👇
Btw huge kudos to @Krzysztof2236 for his work on this feature.
Mobius retweeted
The employee in question
We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI, and we’re grateful to the agents and prosecutors who moved quickly to make an arrest. The FBI’s investigation is ongoing and we will continue to support their work.
At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data. The employee was terminated.
We take protecting our technology, data, and people seriously, and as a matter of national importance. Our security program is designed to protect against a wide range of threat actors, from insider risks to cybercriminals to nation states, and this case is an example of how AI can help address these sophisticated security threats.
Mobius retweeted
Proof that Dots is literally a Codex thread and NOTHING special.
Accidentally found Dot's backend thread and it's just a cloud codex thread.
Nice engineering OpenAI.
Mobius retweeted
the real story is that Tibo is OpenAI's way too cocky secretary of propaganda
Opus 5.5 max is >>> Sol 6.1 max
and actually: Opus 5.5 medium ~ Sol 6.1 max
while Opus 5.5 medium is 1.8x more expensive at API prices, it is also 3x faster per task
on same-priced ChatGPT and Claude plans Opus 5.5 medium completes about 3× as many tasks, because Claude plans include about 5.6× more usage
Mobius retweeted
🚨 I really think OpenAI's FAKE mask is finally over and the REAL face is starting to SHOW UP.
At first, I thought these were just bad decisions.
Now there are too many of them to ignore.
It’s becoming a pattern.
And that’s what worries me.
→ No more Codex resets
→ Our usage is getting nerfed
→ They don't fix anymore, they just look for ways to charge more
→ $200 Pro usage getting cut by 50% and becoming 10x
→ A new $500 tier for heavier usage
→ Astra getting nerfed like hell
→ GPT-6.1 Sol is painfully slow
→ And now Dots, which nobody asked for
One thing I've realized from all this…
They don't fix anymore. They just keep finding ways to charge more.
All we asked for was:
→ Better usage limits
→ Codex actually getting fixed
→ A more reliable experience
Instead, almost every recent change feels like users are being pushed towards paying more while getting less.
That’s what worries me.
Looking at all of this together…
Did “PAY MORE, GET LESS” become the actual OpenAI strategy?
Mobius retweeted
The 3 women of the jury interviewed by the local news felt like the prosecution was a bit too biased & harsh against Lindsay Clancy. It's almost as if they were trying prosecute her for a murder or something...
Mobius retweeted
The anonymous juror said, “there was a lot of big personalities in the room.”
Yup. We know. We just met the 3 big personalities. Exactly what we thought — three pushy, annoying, and emotionally charged women pressuring everyone to see it their way.
And when they didn’t get their way, they accused the male holdout of breaking the law.
I hope this is a wake up call to normies, especially normal men, about how truly problematic most modern women are.
Mobius retweeted
Three babies were brutally strangled to death and this witch thinks the trial was hilarious.



































