Nanbeige4.3-33B's agent scores were run with preserve_thinking=true, so old reasoning stays in context. If your harness strips it, don't count on that 80.9 Terminal-Bench 2.1. I'd test on vs off first. Qwen3.8-27B still wins SWE-Bench Pro, 61.7 to 60.3. x.com/nanbeige/statu…
🚀Excited to announce Nanbeige4.3-33B!
To push the performance ceiling of small agents, our strategy in Nanbeige4.3 starts with building a powerful teacher first.
Nanbeige4.3-33B delivers SOTA agentic performance among open models at ~30B parameters:
* Strong results on agentic coding and cowork benchmarks.
* When integrated into mainstream agent harnesses, it tackles real-world tasks with cost-efficient inference.
* It will serve as a teacher for upcoming smaller Nanbeige models, starting with Nanbeige4.3-3B.
huggingface.co/Nanbeige/Nanbe…
huggingface.co/Nanbeige/Nanbe…
#LLM #Agent
1
0
0
26
If CI is prod, CI secrets are prod secrets. LiteLLM 1.107.0 (Oct 21) shipped a .pth that ran on every Python start, no import needed. A poisoned dep in a job with deploy keys gets them first. Install deps with no creds, mint OIDC tokens for deploy. x.com/championswimme…
At the rate at which people both deploy their own changes to production, and also the rate at which CVEs are dropping which requires patching your systems, the build pipeline and CI systems are not separate from prod in any way anymore.
In fact any and all downtime of build systems are the same as production being down. All software is now ever evolving and ever changing.
The nature of how systems behaved up until a few years back was that in most systems if the service was up and running, the ability to push changes or build it could be down for a while without your service itself suffering from any down time.
Now with agents and the speed of change growing to what it is - losing the ability to deploy is as good as losing the service itself because you have to continuously be updating it.
0
0
0
13
Where fan-out earns its keep for me: GLM orchestrator, Flash workers on narrow lookups, and nothing a worker returns is trusted until compile/test/lint pass on the result. The gate checks the code, so a lossy summary can't sneak through. x.com/dexhorthy/stat…
Subagent overuse and abuse has been a problem since mid 2025.
There is a time and a place and they can be very high leverage, but the vibeslop dopamine casino + twitter hype machine has let the average user down by encouraging more and more parallelism instead of thoughtful context management
0
1
1
29
Chirag Sehra retweeted
Laya can help protect your website from agentic attack jailbreaking. Imagine a swarm of agents attacking your website, and in under 33 ms, it all gets blocked. Also, before the agent is allowed to call a tool (send email, query a database, execute code, etc.), Laya can answer: “Is this action consistent with the original user request and safe under the current policy?” Low confidence or a clear “no” triggers a human review or hard block. All on your browser with the 421 M model on edge!!!
github.com/NandhaKishorM/…
github.com/NandhaKishorM/…
14
38
406
10,532
Prefill is the number for agent use. At ~200 tok/s (per a reply), a 30K-token coding context is ~2.5 min before the first token on every cache miss. On 64GB I'd use it for short-context work like routing or triage. x.com/coldniko/statu…
0
0
0
41
Before putting UUIDv7 in public URLs: the first 48 bits are the creation time in ms, so anyone with the ID can see when the row was made. Fine for internal keys. For user-facing ones I'd keep v7 as the primary key and expose a separate random ID. x.com/arpit_bhayani/…
Databases and systems love it when your IDs are sorted. It makes deep pagination so much more efficient. But if you are relying on UUIDv7 for monotonicity and its timestamp, here is something for you...
First, why do sorted IDs matter? Most databases keep the primary key in a always land on the rightmost leaf page, which stays hot and fills cleanly. Random IDs land on random pages, which means cache misses and page splits.
UUIDs were almost always fully "random", but UUIDv7 puts a 48-bit unix millis in the most significant bits, so IDs sort by time. Within the same millisecond, the order is still arbitrary unless you use a counter. As per RFC 9562, a UUIDv7 implementation can use 12 bits for a counter.
Now, what if the counter rolls over?
With 12 bits, one generator can issue 4096 IDs per millisecond, roughly 4 million per second. Now, imagine a situation where you run out of this (highly unlikely in a typical request-driven service). When it does, the RFC says you can opt for one of two things:
1. Stall until the next millisecond, or
2. Bump the timestamp ahead of the real clock.
So, essentially, we are bumping the timestamp to t + 1 while the system clock still says t. This means that under sustained load, the generated timestamps can run ahead of the real clock. So, you cannot and should not assume that you can use the timestamp from a UUID as your `created_at` timestamp to save a `created_at` column as a micro-optimization.
So, check what your library does on rollover and how many counter bits it uses. Some use 42 or 60 bits, which makes rollover practically unreachable. It is better to build that clarity before you decide to rely on them.
I mean, at the end of the day, sorted IDs are only as sorted as the generator behind them.
0
0
0
16
Laya a month in, on phones and Vercel's gateway. The use I'd watch is agent routing: with calibrated probabilities you set a confidence floor and send only the unsure calls to a bigger model. Curious what floor people land on. x.com/Nandakishorm1/…
After one month of launching Laya decision model. The first open-source version of jev and first ever open-source implimentation of RLCD pipeline. Here are the numbers. And Vega U series are gradually disrupting the market for large context multimodal decision model with custom n without retaining
github.com/NandhaKishorM/…
github.com/NandhaKishorM/…
0
0
0
19
K3.1 Pro API is live now, weights by Nov 4. The blog says cyber data was cut "for the open-weight release." Is the API serving that same checkpoint? If yes, evals you run this week carry over to self-hosting. If not, plan to rerun them on the weights. x.com/Kimi_Moonshot/…
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…
0
0
1
21
What I like here: every method comes with a check that forgetting actually worked, via membership inference and a retrain baseline. Most unlearning write-ups stop at the method. Free, runs on a laptop CPU, opens in Colab. x.com/shreyanbr/stat…
Fresh out of the oven! 🍲
Learn how to make AI forget without paying to retrain from scratch, and how to check that it really did.
I believe this is the first Machine Unlearning course out there, and it's open source and free on GitHub for people to learn from and contribute to.
Hope you enjoy this repo - github.com/Shreyan1/machi…
Every problem follows 5 steps ->
1. Train
2. Show it still remembers
3. Unlearn
4. Show the result
5. Prove it against models that never saw the data
The networks are written in pure NumPy, so you can read every gradient.
1
0
1
45
I ran Tencent's Hy-Memory against 8 hosted models as long-term memory writers. The cheap ones invented 6 facts nobody said. The pricier ones up to 18.
If a model can mutate long-term state, malformed output is a failed write, not a parsing edge case. x.com/i/article/2113…
If a model can mutate long-term state, malformed output is a failed write, not a parsing edge case. x.com/i/article/2113…
10
0
6
517
Chirag Sehra retweeted
Fresh out of the oven! 🍲
Learn how to make AI forget without paying to retrain from scratch, and how to check that it really did.
I believe this is the first Machine Unlearning course out there, and it's open source and free on GitHub for people to learn from and contribute to.
Hope you enjoy this repo - github.com/Shreyan1/machi…
Every problem follows 5 steps ->
1. Train
2. Show it still remembers
3. Unlearn
4. Show the result
5. Prove it against models that never saw the data
The networks are written in pure NumPy, so you can read every gradient.
Learn how to make AI forget without paying to retrain from scratch, and how to check that it really did.
I believe this is the first Machine Unlearning course out there, and it's open source and free on GitHub for people to learn from and contribute to.
Hope you enjoy this repo - github.com/Shreyan1/machi…
Every problem follows 5 steps ->
1. Train
2. Show it still remembers
3. Unlearn
4. Show the result
5. Prove it against models that never saw the data
The networks are written in pure NumPy, so you can read every gradient.
0
1
8
197
Cursor agents can now pick Computer Use on their own when a task needs it, across any app, not just the browser.
First thing I'd test: a separate macOS user with only the Simulator and test accounts signed in, then see what it reaches for. x.com/cursor_ai/stat…
First thing I'd test: a separate macOS user with only the Simulator and test accounts signed in, then see what it reaches for. x.com/cursor_ai/stat…
Cursor can now control your computer.
Agents can drive desktop apps like the iOS Simulator to test the features they build.
0
0
0
14
Happy 1024 & happy weekend, devs.
500M free tokens on ZCode: GLM-5.3 (200M) + GLM-5.3-Flash (300M). 200K packs, while they last.
Claim now. Use Oct 23 15:00 – Oct 26 01:00 UTC.
Update to the latest ZCode, claim in-app, then switch to Start Plan.
zcode.z.ai
0
0
0
8
Ran the spec validator on my GLM skill: 1 failure, a when_to_use field it doesn't allow. Easy fix.
Harder part: the body names 3 ZCode tools and GLM-5.3 models 8 times. Passing won't make that portable.
What breaks when you move a skill? x.com/ClaudeDevs/sta…
Harder part: the body names 3 ZCode tools and GLM-5.3 models 8 times. Passing won't make that portable.
What breaks when you move a skill? x.com/ClaudeDevs/sta…
We're donating Agent Skills to the Agentic AI Foundation (@AgenticAIFdn), the same foundation we donated MCP to.
Since December 2025, 70+ agent platforms have integrated the standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code.
Here's what that means for you:
0
0
0
13
I'd split cost-per-PR by the gate where each run dies. With cheap Flash subagents behind compile/test/lint gates, per-token price looks great, but retries that never clear the gates can eat it. Pass rate per model belongs on the board. x.com/zachlloydtweet…
If you track software factory metrics, don't celebrate the wrong signals.
Bad: LOC written, tokens used
Good: Cost-per-PR, cycle time from conversation start -> merge
A factory leaderboard should take cost into account and champion shipped work, not token spend.
0
0
0
9
If the agent extracts the parts it needs, I'd have it port the library's tests for those parts and run them as a CI gate. Otherwise you've traded a dependency you could audit for code nobody has checked. x.com/TheAhmadOsman/…
Programs you write in the future have to be self-contained. Every surface you touch will be part of that program. Libraries are a supply chain disaster full of vulnerabilities.
Point an agent at a library / API and tell it to extract / rewrite what it needs to get the job done.
0
0
0
10
Check the license before shipping this. Under the MiniMax Community License, commercial use means showing "Built with MiniMax M3.1-Flash" and emailing them a notice. Over $20M revenue needs written approval. x.com/MiniMax_AI/sta…
MiniMax M3.1 Flash weights are now public. 428B parameters total, 23B activated per token.
Agentic performance: A generational leap in agentic tasks, coding, and cybersecurity — 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, and 82.2 on CyberGym.
Efficient by design: MiniMax Sparse Attention and DSpark speculative decoding improve inference efficiency.
Hugging Face:
huggingface.co/MiniMaxAI/Mini…
ModelScope:
modelscope.cn/models/MiniMax…
API: platform.minimax.io
M Plan: minimax.io/m-planMiniMax
MCode for desktop: agent.minimax.io/download
0
0
0
27
Anthropic's Claude Code stat was 89% of dangerous commands caught, so about 1 in 9 got through. I'd keep a plain deny list in front of the classifier for anything you can't undo, like force pushes or deletes outside the workspace. x.com/claudeai/statu…
Starting next week, sessions across Claude web, desktop, and mobile apps will run in auto mode by default on Pro, Max, and Team plans.
Rather than asking you to approve each step, an automated check blocks risky actions. Switch back to manual anytime.
1
0
3
46
Worth knowing from the methodology: if a provider refuses and the agent falls back to another model, that attempt is scored like any other. So a row can include work a different model finished. I'd check the fallback rate before comparing rows. x.com/ArtificialAnly…
Announcing Artificial Analysis Coding Agent Index v2.0: FrontierCode v1.1 from Cognition replaces DeepSWE, SWE-Atlas-QnA grading is revised, and Claude Opus 5.5 in Claude Code leads at 61.7
Today we are launching v2.0 of the Coding Agent Index. The Index evaluates models nt harnesses, now across 290 tasks covering software engineering, terminal use, and repository Q&A, with the three benchmarks weighted equally.
Changelog (Index v1.5 → Index v2.0):
→ FrontierCode v1.1 replaces DeepSWE v1.1. FrontierCode is a benchmark from @cognition with a private test set: 100 software engineering tasks on real open-source repositories, graded on whether a maintainer would merge the change. An attempt that fails a blocking check scores zero, and one that passes them all scores between 70% and 100% depending on the non-blocking checks it meets. We also scan each trajectory with a modified version of Cognition's rule-based scanner, and any attempt flagged for accessing solution material scores zero.
→ Revised SWE-Atlas-QnA grading, now SWE-Atlas-QnA-AA. Scale AI's 124 repository Q&A tasks are unchanged. We reworded 127 of the 1,306 rubric criteria for more consistent verdicts and moved to GPT-6 Luna (xhigh) as the judge, with five votes per criterion. The dataset and judge configuration are published on Hugging Face.
Key results:
➤ Claude Opus 5.5 leads the Index: Claude Code with Claude Opus 5.5 (max) scores 61.7, followed by Claude Code with Claude Sonnet 5.5 (max) at 59.3, Muse Code with Muse Wonder Full (max) at 58.9 and Claude Code with Claude Fable 5.1 (max with fallback) at 57.6.
➤ FrontierCode v1.1 is where Opus 5.5 pulls ahead: Opus 5.5 scores 56.0% on FrontierCode, ahead of Codex with DeepSeek V4.1 Flash (max) at 52.6%, Codex with GPT-6.1 Sol (max) at 49.4% and Codex with GPT-6 Astra (max) at 49.3%. Below the top, FrontierCode scores are tightly packed: 15 of the 24 configurations on the Index score between 43% and 49%.
➤ SWE-Atlas-QnA-AA grading is stricter and more consistent: across the 21 configurations evaluated on both Index versions, the mean score falls 1.7 points, from 58.8% to 57.1%, with 14 of 21 moving down. Individual configurations move by up to 7 points: Muse Code with Muse Spark Open (xhigh) falls 7.0, Kimi K3 in Kimi Code CLI and Codex with GPT-6 Luna (max) 6.2, and Muse Code with Muse Wonder (max) 5.6, while Codex with GPT-6.1 Sol (max) gains 2.1.
➤ Four configurations sit on the cost frontier: Codex with DeepSeek V4.1 Flash (max) scores 47.3 at $0.39 per task, Codex with GPT-6.1 Sol (max) 54.1 at $1.50, Codex with GPT-6 Astra (max) 54.7 at $7.04 and Claude Code with Claude Opus 5.5 (max) 61.7 at $11.42. Costs are pay-per-token API costs, not subscription prices.
We will keep updating the Coding Agent Index incrementally, with more evaluations and agents to come.
0
0
0
8
What I'd copy from METR's writeup if you run agents behind a monitor:
- run it every N steps, not at turn end
- reconcile inference logs against monitor logs
- feed it the same serialized context the agent saw
- no client-set header should turn it off x.com/OpenAI/status/…
- run it every N steps, not at turn end
- reconcile inference logs against monitor logs
- feed it the same serialized context the agent saw
- no client-set header should turn it off x.com/OpenAI/status/…
This summer, we invited METR and Redwood Research to pilot an embedded evaluation of our internal monitoring system, providing deep access to our infrastructure over three weeks.
Their findings helped strengthen safeguards as we expanded the system and will help shape future research collaborations with METR and other third party evaluators.
0
0
0
7