پن کیا گیا ٹویٹ
If what you're working on is not important and it's not likely to lead to important things, why are you working on it? - Richard Hamming
English
Uday Bhaskar
454 posts
Uday Bhaskar
@BhaskarSteve
Techno optimist. Prev: @iiit_hyderabad
شامل ہوئے Haziran 2020
1.2K فالونگ172 فالوورز
English
Currently if you know you have to chase a low total, teams are usually pacing the innings to finish it in 18 overs and sometimes but very rarely it goes to last over and it's somewhat fun. With this incentive teams will try to finish in 12 overs for the extra point and if they collapse in the process, they will change plans and aim for the win instead of the extra points. Same goes while defending, teams will keep attacking if they can finish it early or go back to defensive lines if its going wrong. Both ways its more entertaining.
English
@EducatedMoron @lyrical_guy20 It just gives added incentive to finish strong, which can go both ways and in both ways the audience win because its more entertaining. I agree that 2:1 bonus points (2 points for win, 1 bonus) will make it unfair. But 4:1 bonus points is not as bad and more entertaining.
English
What do you think NRR means and supposed to take care of?
Aakash Chopra@cricketaakash
I wish IPL adopts the Bonus Point option for a massive win…
English
@EducatedMoron @lyrical_guy20 SA20 uses 4:1 bonus points and I think it’s working really well
English
@lyrical_guy20 Yes it would makes matches more interesting. But bonus point is double rewarding a team, first they get a big jump in NRR & then get a bonus point too.
NRR alone keeps things very fair. Two teams won 7 games and now lets see which won them easily. Simple.
English
Everyone is giving genuine answers but I think they’re hinting that they’re working on pure RL approaches without pretraining. I remember Jerry went on Matt Turck after the Sutton Dwarkesh interview and he was asked about pure RL without pretraining and he just said we’re doing very serious RL work but we still need pretraining. I think he also later mentioned he was leaving OpenAI because he didn’t have enough freedom and compute to pursue some serious risky research direction. It’s a long shot but it does add up.
Maybe bitter lesson is gonna bite us hard again.
English
@ChangJonathanC @thsottiaux It happens only after a compaction, mostly because internally its considered a new turn
English
soon, we will ship models faster than the average rollout and you will need to update your model mid rollout
Peter Steinberger 🦞@steipete
I love this.
English
@allgarbled Yes. It’s counterintuitive but extended thinking actually helps now because it can code-gen “helper images” in the CoT to use as multimodal input for the final generation.
Pro is clearly a better image generator than even Thinking Heavy though.
English
Can anyone explain to me how it can do this but it can’t consistently generate a nine sided polygon?
Riley Goodside@goodside
ChatGPT Images 2.0 generates a game die but instead of numbers it has working QR codes for each of their Wikipedia articles
English
RLHF, but it's just good code vs shit code
What's the best source to train a RM for this?
English
@eliebakouch @Yulun_Du I found this very useful. x.com/KarelDoostrlnc…
As peter says "Just talk to it"
Karel@KarelDoostrlnck
English
Kimi K2.6 helped us rewrite kernels; it worked like a charm :) #diff-00a0cb8587fadafec814d381a69c346fb0274279c82e25ea709251603b26f7b4" target="_blank" rel="nofollow noopener">github.com/fla-org/flash-…
Kimi.ai@Kimi_Moonshot
Meet Kimi K2.6: Advancing Open-Source Coding 🔹Open-source SOTA on HLE w/ tools (54.0), SWE-Bench Pro (58.6), SWE-bench Multilingual (76.7), BrowseComp (83.2), Toolathlon (50.0), Charxiv w/ python(86.7), Math Vision w/ python (93.2) What's new: 🔹Long-horizon coding - 4,000+ tool calls, over 12 hours of continuous execution, with generalization across languages (Rust, Go, Python) and tasks (frontend, devops, perf optimization). 🔹Motion-rich frontend - Videos in hero sections, WebGL shaders, GSAP + Framer Motion, Three.js 3D. 🔹Agent Swarms, elevated - 300 parallel sub-agents × 4,000 steps per run (up from K2.5's 100 / 1,500). One prompt, 100+ files. 🔹Proactive Agents - K2.6 model powers OpenClaw, Hermes Agent, etc for 24/7 autonomous ops. 🔹Claw Groups (research preview) - bring your own agents, command your friends', bots & humans in the loop. - K2.6 is now live on kimi.com in chat mode and agent mode. For production-grade coding, pair K2.6 with Kimi Code: kimi.com/code - 🔗 API: platform.moonshot.ai 🔗 Tech blog: kimi.com/blog/kimi-k2-6 🔗 Weights & code: huggingface.co/moonshotai/Kim…
English
Very cool work, Congratulations!
If you're training domain specific experts, why not consider distillation based approach like MOPD instead of a modular approach. Training remains on policy and sample efficient with dense rewards, hence minimal forgetting. Specialist architecture is also flexible across size, architecture and training (except for same tokenizer, which is also relaxable) and the base model architecture also remains same which is convenient.
English
How do you add new capabilities to a fully post-trained language model, without retraining from scratch, or losing what it already knows?
We're excited to introduce Branch-Adapt-Route (BAR): train independent experts, merge them into an MoE, and upgrade them as needed.
Ai2@allen_ai
Last year, we introduced FlexOlmo, a novel way to train parts of a model independently then combine them later. BAR builds on that idea for a harder problem: how to keep improving a model without having to retrain each time. 🧵
English
I have many, many thoughts catching up (bc they released after my bedtime). M2.5 has ran my Claw since Claw was first a thing. However, I will look at if quantized GLM 5.1 > Minimax over the next few weeks and change some workflows.
English
I’m honestly unsure if I’ll make a model card for this.
I also can’t really run this model at home given this restriction.
There are many good alternatives now, at least.
Florian Brand@xeophon
wow, they did a non-commercial license... M2: Display the name if >30M revenue / 100M users M2.1: Display the name M2.5: Acceptable use policy M2.7: Non-Commercial license
English
Going under the radar but quietly building an amazing product, @interaction is a generational company in the making. Their attention to detail, how they handle their business and the quality of the product they are offering will not go unnoticed for long.
English
@pashmerepat I’m codex monothread pilled too until I see the rate limits disappear
English
Take the monothread pill 💊
Nick@nickbaumann_
So much coding agent design is built on the assumption that breaching context windows and compacting context yields progressively worse results (/newtask, many threads, etc) When you drop this assumption, the product direction it opens up is very exciting :)
English
@TheZachMueller I also noticed we can’t search for text in bio in the people section. This was very useful to search for company affiliates
English
So… we’ve removed the single most useful thing on this app for networking, the ability to see mutuals?? Am I getting this right?
English
Congratulations on 50 glorious years to the greatest company every built in my lifetime
English
@rohanvarma @dkundel More tokens without degrading performance? Yes please
English
If we made /slow mode in Codex, would you use it? What for?
(Slower inference at a cheaper cost)
English