Wing Chan

@Designarena Without having vision too?

@SubarcticRec @VadimStrizheus What quantisation? There should be community pools where you spread the cost within a private pool

@VadimStrizheus Deploy the model to RunPod and you control the limits with your wallet.

What’s all this hype about GLM 5.2 being AGI?? I’m constantly getting rate limits and errors.

Day 2.

@chetaslua Lite version. Speed is good

🚨 Google New Image Model > Instant-ramen (successor of nano-banana) Ramen is cooked time to serve soon , we will share results as soon as we get hands on it 😉

@reach_vb So impressed. When codex mobile first came out I was pretty worried due to the lag /connection issues, it felt really unfinished. But now it is rock solid. Congrats on finding the path through, probably helped by a mythical beast of a model yet to be unleashed

Codex Mobile updates: - browse workspace files and link paths into prompts - pick a workspace folder when starting a new thread - expand or collapse all diffs while reviewing changes - approve MCP actions for one chat or across chats - LaTeX rendering in Codex messages and plans - clearer status for running threads, queued prompts, side chats, and subagents - better pairing, onboarding, reconnects, host refresh, and thread performance - improved Codex profile sharing, activity history, settings, transcript layout, and assistant actions - smoother /goal workflows from mobile - fixes for stuck swipes, duplicate messages, subagent rows, and misleading connection errors

@MicahBerkley @SubarcticRec I'm excited to move to a benchmark that is the cost to finish a project to MVP launch state. And I think for most cases the overall cost will be dominated by the human time spent and not the token cost, unless you are using fable etc. Nothing great can really be one shot.

@SubarcticRec valid... .but still there's a nest of first to market guys that I'm used to seeing on my timeline.

I haven't seen anything dope designed by GLM-5-2 yet. Would have thought to have seen a flooded timeline... #KanyeShrug

@SubarcticRec Simple text it does great. Struggles if more than a few words probably model size issue

Have not even tested before, but Flux.2 Flash and ZiB both do correct text.

Day 1 of asking @zoink if he will consider adding RF2.5 to Figma Weavy and Figma. Especially now that I know his email inbox is not the right route.

@ArtificialAnlys Great to see the care and attention in keeping these single number measures refreshed. Thanks for sharing

Following up on our Intelligence Index v4.1 release yesterday, in the video below, Daniel from our team shares a short overview of what's changed: 1. Three upgraded evaluations: Terminal-Bench 2.1, τ³-Bench Banking and GDPval-AA v2 2. Cost, time, and tokens per task: Understand the cost, time, and tokens of tasks across our Index and for individual evals, and how these trade off against Intelligence 3. Cached input token reporting: We now report the amount of cached tokens a particular model uses and how this influences cost

BREAKING: Riverflow Pro 2.5, a reasoning model by @riverflow_ai that calls a mix of proprietary and open diffusion models, has scored 1st on Image Arena (Models + Routers), 1st on Graphic Design Arena, and 1st in Image Edit (Models + Routers). Riverflow Pro 2.5 averages 10 Elo points above GPT Image 2 from @OpenAI in Image, Image Editing, and Graphic Design. It also establishes Pareto frontiers across Image, Image Editing, and Graphic Design in Preference vs. Speed. Congratulations to the @riverflow_ai team on the launch!

@grx_xce @MistralAI Lucky they aren't drawing cat memes yet phew

BREAKING: Le Chaton Fat has fully saturated our benchmark. We are at a loss for words. In response, we are retiring Design Arena. Congratulations to the @MistralAI team, and thanks for putting us on vacation.

This is a high difficulty problem to solve. It's another form of alignment. To help me make my code secure and good, I need the model to know how to break it. To stop it from attacking others, I need the model to resist me. But it depends who is asking and that's non trivial. Think about the surface area, with agents, context, tool calling, harnesses etc. How do you verify? How do I trust the result of a tool call? We end up in the same place, training the model to make moral (i.e. non trivial non binary grey area) decisions based on some imperfect principles and imperfect data. data. The hardest part of all this is that since model training now exists across a broad array of actors, there always exists the incentive to offer a version slightly more permissive and greedy.

@Salmaaboukarr Have you tried minimax m3

this claude fable 5 drama made me realise how important it is to have a backup plan and not rely on these labs! time to go back to LOCAL MODELS + a v good harness the models i use -qwen -kimi -ds

@grx_xce Alive? Good

He still hasn’t woken up yet

@nicdunz Automation?

1. i did not get the option to save this reset and not use it 2. why is my usage already a few percents down when i havent even used it yet?

@jackfriks @postbridge_ Great journey!

can't believe this was only 18 months and 8 weeks ago...

@Sauers_ Fat if true

Big if true

@Ronanchamberss @KanishkaNarayan @etnshow Cover more stories outside of London too please! Bring the whole of the UK for the journey

After speaking with UK AI Minister @KanishkaNarayan today on @etnshow, I have it on good authority that, we are indeed, so back. Best, R

The car is not really the deliverable though

@petergostev Should get 10x points for loc reduction if all tests pass. Code golfing is fun but the principle applies. Verbosity is a feature of bad compression.

if they are counting loc i'm screwed