Pinned Tweet
GLM-5.3 Flash took 14.5% less time to finish my task, even though DeepSeek generated more tokens per second.
I ran both models using Codex CLI on my M3 Ultra:
• DeepSeek V4.1 Flash : 79.4 tok/s, finished in 69.8 seconds
• GLM-5.3 Flash : 61.5 tok/s, finished in 59.7 seconds
DeepSeek's answer was 48% longer : 2,991 VS 2,016 tokens
If you've tried DeepSeek V4.1 Flash 1017, has it fixed the issue of giving longer answers than needed?
Damn. T3 Code makes my coding subscriptions feel like one app.
I can switch between Codex, Claude, Grok and my DeepSeek harness, and my context carries over automatically.
It opens instantly on my phone, too. If you’ve used Codex on mobile, you know the startup wait
I can finally track usage across my subscriptions in one place : ~3.4B tokens burned so far
@theo cooked with this one.
Uncensored GLM 5.3 Flash on my M3 Ultra has been doing the security audits Codex kept refusing
It's done every task I've asked so far. Not a single "I can't help with that."
I tested the DeepSeek Search API, but I still prefer Codex and Claude Code's answers.
I was looking up new AI papers and model benchmarks. Their built-in web search tool answered my questions more directly, so I don't see a reason to add a separate search API for those searches yet.
I can see it being useful if your agent doesn't have a search tool connected.
Server-side search through DeepSeek's API is free until 1 Nov 3AM AEDT.
🔍 Search & Fetch APIs are now open to everyone!
Powered by our in-house search engine, serving hundreds of millions of requests daily worldwide.
đź’° Each endpoint: $0.75 per 1,000 calls.
📖 Docs: api-docs.deepseek.com/guides…
đź“Ś Search & Fetch calls via server-side tools in DeepSeek Harness, Claude Code, and Codex are currently free for DeepSeek API users. Billing starts Oct 31, 2026, at 16:00 UTC.
3/4
This may be the strongest endorsement of DeepSeek V4.1 Flash capabilities.
Dario wants to close every loophole around chips, compute access and distillation.
My read: Chinese AI labs is less than 3–6 months away from overtaking the US, much closer than earlier estimates.
Slowing their competitors would also let OpenAI and Anthropic keep cutting our usage limits whenever they want.
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must…
Codex reset incoming in about 3h 30mins (12AM PST)
Hi Astra users. A reset and a quick update on quality issues that have been posted around.
Working with some of you, we have found and fixed the following issues:
- Some skills written for previous models were triggering too often or preventing the model from checking its work.
- An opt-in context management experiment that could cause early stops or replies to older messages. We've disabled it. Our rough estimate is that 4-5k users were affected by this experiment.
- We've also removed some badly configured engines that resulted in a measured quality degradation for a long tail of traffic flowing through them.
We’ve also made some more minor improvements and things should feel significantly better across the board. More consistent follow-through, better tracking of your latest message, and better checks on the work as it’s going through the motions.
The examples posted and all the users who worked directly with us were incredibly useful in helping fix things quickly. Always grateful for this incredible community.
And of course, a reset is also landing by midnight today.
Prediction: By Christmas, Chinese AI labs will release open-weight models with Fable 5-class intelligence that you can run locally.
DeepSeek Flash model is now approaching 5.6 SOL-level intelligence. 🤯
Early V4.1 Flash results:
• Native vision support
• Reported speeds of around 300 tok/s
Token price /1M (input / output)
V4.1 Flash : $0.30 / $4
5.6 Sol : $1.20 / $20
Deepseek is 13Ă— cheaper on input, 16Ă— on output.
DeepSeek even routed V4 Pro traffic to V4.1 Flash.
Can’t wait to run this beast locally.
🧠Asymmetric architecture. More intelligence, less cost.
🔹 552B-parameter MoE.
🔹 New Causal Encoder–Decoder architecture: just 8B active parameters for input, 16B for output.
🔹 New pre-training methods + larger-scale RL post-training deliver benchmark results ahead of flagship models, including DeepSeek-V4-Pro.
2/6
Most people think the white-collar job apocalypse already started. OpenAI who wrote that narrative now describes a gradual curve.
The old pitch was simple: AI will replace huge amounts of white-collar work.
For CEOs, the appeal was obvious
↓ labour costs
↑ productivity
an edge over competitors
Waiting felt risky.
And it worked : Anthropic says business customers drive ~80% of its revenue.
Now Sam Altman says AI capability can improve exponentially while society changes gradually.
While models can improve rapidly. Companies need years to change workflows, retrain staff and manage risk. Large institutions often take even longer.
We’re still early, learning AI now isn’t too late.
Something feels off with Claude lately
Opus 5 kept giving me fixes that didn’t work. It took multiple attempts to solve a PySpark error that 5.6 Sol fixed on the first try.
Claude used to be the best. Pro has been cancelled for now.
Waiting for M5 Ultra could mean paying more for the same memory capacity
Apple paid more for memory in March than December,
significantly more in June than March,
and they expects another increase in September
For Local AI, buy in this order :
256GB unified memory Mac Studio > 2 DGX Spark > wait for M5 Ultra
If current hardware handles your workload, being 1 generation behind might be smarter than gambling on next-gen pricing
Apple's earning call 👇
Stop treating shared AI chats like private links
I downloaded a public archive of 755 Claude chats. 1 included a CSV I couldn’t open, but Claude’s analysis of it was still readable.
If you’ve shared anything sensitive, find and delete the conversation to revoke it
Paul Ho retweeted
The negative AI jobs outcome just continues to not be happening as some predicted. A large portion of enterprises I talk with -across industries- are still hiring, just with a tilt in the kind of roles they’re going after.
They’re finding that AI is letting them do more, and Jevons paradox is actually playing out. They’re hiring engineers to go after problems they couldn’t tackle before. They’re hiring in sales because they can go deeper in client relationships now with the help of AI. They’re hiring in internal FDEs to help them deploy AI. And so on.
Anyone using AI merely to cut costs eventually just gets outcompeted by companies that use AI to better serve their customers and drive more breakthroughs in their business.
Could this trend change at some point? Sure. But for now this appears to be the trajectory we’re on.
