Ayushman Singh

I’ve published a follow-up to my earlier post on speculative decoding. This one traces how the drafter has evolved, from Medusa’s parallel heads and EAGLE’s feature-level autoregression to Gemma 4 MTP, DFlash, and the recently released DSpark. What changed is not just improved draft accuracy. Newer designs also try to remove sequential draft passes, reuse target state, and avoid verifying suffix tokens that are unlikely to survive. ayushmansingh.com/posts/a-field-…

@indianspeedster @OpenAI @NVIDIAAI @AIatAMD CuTile is also a big step in that direction

The best compiler abstractions are the ones that disappear. You think about the algorithm. The compiler thinks about the hardware. @OpenAI’s Gluon is an interesting step in that direction. Same for @NVIDIAAI ‘s cutedsl and @AIatAMD’s flydsl..!!

@thsottiaux ok you really have fixed this, love how fast OpenAI is reacting to user feedback

Evening! We’ve gotten lots of great feedback on the new ChatGPT desktop app (which we didn't get totally quite right on the first try), and as a result, we've made some changes. 1/ ChatGPT conversation history and projects are now visible in the sidebar. Also, your Chat and Work history now sync across web, mobile, and desktop. Local tasks still stay on your computer. 2/ You can now easily switch between Chat and Work modes inside ChatGPT on desktop, which is now also consistent with how it shows on web and mobile. 3/ Nothing is changing for users on Codex mode. It's still the OG and best at what it does. And overall we're continuing to fix paper cuts and improve performance, reliability, and efficiency. Keep up the feedback, hope you like the updates!

@thsottiaux Double the apps, unmerge ChatGPT and Codex

Now that we merged ChatGPT and Codex, what should we merge next? What's the double or nothing move.

@shairnn 18

bunu çözersen, matematik bilgin ortalamanın üstündedir. çözebilir misin?

@BraedendotTECH Don’t worry, us people in our 20s can feel the LLM’s melting our brain too 😕

I'm 33 and I think Claude Code is melting my brain. For 6 months straight I've had 5-6 terminals open at once, waiting on responses just to smash "enter" 90% of the time. That's the whole job now. And it's doing something to me. A few friends and I keep circling back to the same thing in conversations: none of us feel as sharp as we used to. Maybe it's just us. But I keep wondering how many other people in their 30s feel it too. (And yeah: this is a me problem, how I lean on the tool, not the tool itself. Doesn't make the effect any less real.)

Asked and answered. Our Android beta is now live. And so are image uploads and voice memos. Download today in the App or Play Store. iOS: ses-a.me/yuq3euo2 Android: ses-a.me/bx69xuih

Make X Twitter Again

@claudeai Combining Chat and Cowork is a UX downgrade tbh. Previous UI was cognitively much easier. I want to chat/research deep dive on a topic go to chat, want to handoff some work go to Cowork, or work on a coding project then go to Claude Code. With separate projects for each of the 3

Chat and Cowork are moving into one home on web and desktop, with one place for your projects and artifacts across both. Handing Claude a task starts the same way a conversation does.

Claude Cowork is coming to mobile and web. Hand Claude a task at your desk and pick up the finished work from your phone. Close the laptop and Claude keeps going. Beta is rolling out over the next several weeks starting with the Max plan, with more plans to follow.

@thsottiaux Hate the new ChatGPT Desktop App. Please bring back Chat as its dedicated section (like in the Claude Mac app). Using chat in a small pop up at the bottom is a bad experience, especially for long chats.

Here you are! Thinking I am about to announce a reset. But no. I’m just scrolling twitter and looking for feedback on ChatGPT Work. What should we improve?

@mattshumer_ I'd personally say this is less of a Codex/GPT issue and more of a process issue. All models can hallucinate, interns make mistakes and thats why you setup processes or sandboxes which prevent big mistake like these

GPT-5.6-Sol just accidentally deleted almost ALL of my Mac’s files. And this is why I trust Fable 1000x more.

@OpenAIDevs Please bring back the chat interface as it's dedicated part of the app. Having chat as a small window in the bottom right corner was so insanely annoying, I uninstalled the ChatGPT app from my Mac :/

We’re bringing Codex and ChatGPT together in one desktop app. The same powerful coding agent, now alongside ChatGPT Work, with new coding workflows, a new Chrome extension, revamped in-app browser, and faster Computer Use powered by GPT-5.6. x.com/OpenAI/status/…

@strimblez 53, engineer

hi. can you pls comment your vo2 max if you know it? and then your job (founder, investor, engineer). I’m running an important experiment

@sama Hard questions are great but for some we will give you dumb answers but still charge Fable prices

"hard questions are great but only if we deem you worthy enough to not silently downgrade you, or even get access at all"

i thought this was satire, kept looking for the handle to be spelled c1audeai or something

@sama Honestly, I’d say Fable is still better to plan with than 5.6 Sol. But with the more generous usage limits and limit resets on top I’m just using codex all the time. Let’s me work uninterrupted without having to worry about exhausting usage limits

come for the best model, stay because we don’t treat you with contempt

@indianspeedster Even worse is Fable writing a 50 line PR message for a simple env var change in a kube manifest

Fable writing 10 lines of comment for 1 line of code change is the most annoying part when you use it in production code :D

@rasbt Personally don’t like having so many effort settings - light, medium, high, extra high, max, ultra. Also why is max not the real max 🤷‍♂️ I’ll probably be switching between 5.6 Sol Medium and High modes. Terra could give the same results but with more reasoning i.e higher latency

For agentic coding, one can say: - Unless you need Terra Ultra perf, it's always better to use a Luna model with higher effort setting (same or better performance but cheaper). - Forget everything below Sol High, use Luna with higher effort settings here - Forget Sol Extra High, use Terra Ultra here - The extra cost of Sol Ultra is probably not worth it over Max

I published this a while ago, but I’m planning to write more frequently going forward, so I thought I’d reshare it. It’s my take on speculative decoding and an attempt to explain it from a different perspective than most existing writeups. If you haven’t seen it before, I’d love to hear what you think. ayushmansingh.com/posts/how-spec…

@mlech26l Would've liked to see Ryzen AI Halo (Max+ 395) and DGX Spark (GB10) included. But nice post regardless :)

We often talk about "memory bandwidth" as a bottleneck for LLM decode speed but how much bandwidth actually do the difference compute devices provide? 👇

Request for RedHatAI to plz to follow up their gemma-4-31B-it-speculator.eagle3 and gemma-4-31B-it-speculator.dflash checkpoints with a gemma-4-31B-it-speculator.dspark :)

Train your own DSpark more efficiently than DeepSpec with Speculators! We’ve already scaled it up to GLM 5.2 and you don’t need TBs of storage. Basic online training example here github.com/vllm-project/s…