Applied AI and Inference Previously @HackerRank, @Automattic /WordPress
Spain
Joined July 2013
- Tweets475
- Following485
- Followers655
- Likes3.2K
The fk, always own harness then these rules.
ANTHROPIC IS CRACKING DOWN ON VIDEO GAME REWRITES 🚨
Claude Code picked up a new rule around Oct 27, sent from Anthropic’s servers under a flag called "tengu_heron_brook". There’s no setting to turn it off.
The full rule: “Do build original games, games in a genre, and games from the user’s own designs and assets. Don’t build a recreation of a specific existing game, even if it is renamed or called original.”
Is this the end of game rewrites, modding and everything in between?
Abhishek Gahlot retweeted
MiniMax M3.1 Flash: Space Performance + Bunny Speed
🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.
API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.
Model weights will arrive on Oct 23.
API: platform.minimax.io
M Plan: minimax.io/m-plan
MiniMax Code for desktop: agent.minimax.io/download
Pulsar: Qwen3.8-27B on M5 Max.
Writing code: 206 tok/s peak and 148 tok/s decode
Editing code: 445 tok/s peak, 321 tok/s average
Repo github.com/loopai-hq/pulsar
Built on Splash
Comparing lithos speed with own project here github.com/loopai-hq/fastker…
In my test runs megakernels didn't give much speed on mac. Results soon.
Comparing lithos speed with own project here github.com/loopai-hq/fastker…
In my test runs megakernels didn't give much speed on mac. Results soon.
We’re open-sourcing lithos-metal 🚀
Megakernels + DSpark speculative decoding run Qwen3.8-27B at 200+ tokens/s/user peak on one @Apple M5 Max.
Ultra-fast inference on your laptop. Try it with any coding agent in one command.
Code: github.com/lithos-ai/lithos-…
Tech blog: lithosai.com/blog/lithos-met…
I am looking for AI interns for @loopai_hq. DM me if you have done exceptional work.
Great perks, salary and convertible to full time. Share me your work.
Potentially unlimited token budget too if you got the skills.
One thing liking in dots is at least the subscriptions isn't confusing. which exactly to buy is it grok, cursor and how to link etc.
Abhishek Gahlot retweeted
The Head of Engineering when he sees PMs pushing vibe slop to production but also has a mortgage to pay
Local board where Claude Code, Codex and OpenCode report their tasks.
nothing fancy something i am trying to avoid losing track of which one is doing what.
github.com/abhishekgahlot2/m…
OpenJev (Loop) is out
huggingface.co/openjev
Four builds are up:
openjev/openjev, 16 bit
openjev/openjev-FP8
openjev/openjev-MLX, 8bit
openjev/openjev-MLX-4bit, 15 GB - For Apple Silicon (runs on a mac)
Benchmarks:
10K questions, same options, same order
Jev 85.4%
OpenJev 84.0%
On H100 approx time is 150-250ms
It definitely wont be as fast but typesafe isnt probably serving it on consumer hardware.
Why I built it. A lot of people are training small models and claiming they beat Jev. Better, faster. Not sure they realize it is a general one. I ran the 7B and 9B fine tuned ones over thousands of questions and compared them to Jev. They perform badly. If I only needed fast, I could train my own BERT. The point is it has to be fast and it has to be a general classifier at the same time.
The model that is small and smart enough for that, as of September 2026, is probably Qwen 3.8 27B.
All the people who are training or claiming victory over Jev that they are better, faster, not sure they are realizing that it's a general one.
And no one even come close, like all the fine tuned one.
And even the 9B, 7B parameter ones when I try to run over like thousands of questions and compare it to Jev they perform badly and some even performed worse than the base version that they fine tuned on and if I have to make it fast I can probably train my own BERT.
But the point it has to be faster and it has to be a general classifier sort of and the model that actually comes in my mind that is small enough and is smart enough as of now in September 2026 is probably Qwen 3.8 27b parameter one and the one that is probably second one is the Qwen 3.6 35B A3B but it's probably old and not as smart so that's not my first choice.
Qwen 3.8 is probably the one I would choose. its still a very good open model that can run fast in a single pass, perform as a general classifier, and maybe come close to Jev.
Abhishek Gahlot retweeted
I am done with this shit. It is over. The state of engineering right now is horrible. It has been half a month since I started a new role at a big company. Nobody knows anything here. The specs, code, tests, PRDs, tickets, resolution of those tickets, reports, etc., everything is made by Claude Code. Nobody on my team likes this. They are being forced to ship as much as they can. I have heard multiple times from higher management that pushing code is not a bottleneck, so why are we slow? People are working 12 to 13 hours a day just to press enter. Nobody is reading anything. Humans in corporate are doing nothing on their own. Everyone, literally everyone, from an L1 to an L7 engineer here is doing the same thing. Talk to Claude. There is no sense of victory. Nobody is resolving bugs. In reality, nobody is thinking anymore. Everything is done by LLMs. It is so soul-sucking. I would not mind it, to be honest, if we were at least given the time to check out the code and see what is going where. But no, the goal is to just ship. No matter what happens.