Pinned Tweet
I was asked whether my agents could help grow sales 👀
Here’s a small Agent Interface demo exploring that question for a candle shop: research, a plan, then the team gets to work.
What business problem would you give your agents?
agentinterface.dev
Congrats to Anthropic on crossing $2 trillion at the IPO open.
One free Claude usage reset per trillion feels only fair. Drop our 2 resets in settings please.
nitter.cf/AnthropicAI/status/212…
4
321
Tested Gemini 4 Neon by having it build a runner game.
The benchmark table puts it right next to Opus 5.5 and GPT-6.2 Sol, though in practice it definitely doesn't feel close to Opus 5.5.
The game itself came out interesting, except I had to record this clip manually because Neon couldn't figure out how to capture its own demo video (something even Sonnet and Haiku handled on their own).
Curious to see how it holds up on actual coding tasks. Anyone else testing it today?
4
5
326
Anyone else feel weirdly indifferent about GPT-6.2 Astra and Sol?
I wired them into my pipeline and they work, yet I haven't felt any urge to build demos or play around with them.
Honestly think the lukewarm reaction has less to do with the models and more to do with hype fatigue. When you tease massive drops, miss the moment you set up, and talk about security safeguards instead, people stop dropping everything for your launches.
Still need to stress test them properly. How are they actually performing for those of you pushing them today?
1
4
105
Anthropic aiming for a $1.4T IPO explains why the "OpenAI has huge models coming" whispers started right on cue.
@grok do you think OpenAI drops a major model before Anthropic's IPO to protect its valuation?
Curious where this stands right now. @grok how many people have signed this petition so far, and what's the realistic probability Jensen actually returns as CEO?
standwithjensen.com/
1
3
111
Every AI chat app needs an "I told you so" button.
When you share a non consensus idea, agents push back with the safe textbook answer. Arguing past the median opinion over and over gets exhausting.
Imagine logging the disagreement, coming back when your prediction or design works out, and updating the score so the agent learns where to trust your judgment.
Where do you find yourself fighting default model pushback the most?
Why did OpenAI go quiet right before Anthropic's IPO?
Look at the incentives. Anthropic is about to set the first public price tag for a pure frontier lab. If OpenAI drops a huge release now and hurts that debut, they cap their own future valuation too. Let Anthropic price high, ship two weeks later, and tell investors what number one is worth.
Meanwhile Google keeps shipping all week because a rival IPO does nothing for them.
How often do you run your coding agents in YOLO mode?
Permission prompts kill momentum, yet skipping them without isolation always feels one bad command away from a mess. I am working on a sandboxed mode in Agent Interface to get rid of prompts safely, and I am curious what stops you from running skip permissions all the time.
2
73
Models clear the familiar ground so you can spend your energy on the frontier.
Anything resembling past work gets handled fast now. The unsolved designs and first of their kind problems still need a human present, because there is no prior pattern to replay.
What kind of work on your plate right now still feels completely out of reach for agents?
1
48
Models act like a developer with ten thousand years of experience on familiar problems.
They move fast and sound certain whenever the task resembles something they have seen before. We call it intelligence, though most days it is massive, well indexed experience doing the heavy lifting.
What was the sharpest reminder you got recently that a model was relying on memory over reasoning?
Naming the skill that keeps humans valuable is the first step to automating it.
Once we can draw a clear line around good and bad examples, a model can learn the pattern.
Our real edge lives in the parts of the work we still do not know how to describe.
Where in your projects do you rely on judgment you could not explain in a spec?
1
2
75
Everyone has access to the exact same frontier models, yet the outputs could not look more different.
Picture several people getting one hour with a world class expert. What you walk away with comes down to what you bring, what you ask, and what you build after.
You do not need a buzzword like taste for that.
What do you bring into a session that changes what the model gives back?
2
3
77
What is the next word we will invent after "taste" gets automated?
First we called the human edge prompt engineering, and now we call it taste.
Models are already picking up better taste with every release. What we bring is real, it just does not fit inside a single named skill.
4
2
86
Read the Stand with Jensen letter and consider adding your name. #standwithjensen #bringjensenback standwithjensen.com/
1
5
172
Around 2018, headlines kept hyping AI models that wrote Shakespeare sonnets and half-gibberish books. I remember thinking it was completely pointless.
I was working on translation models back then, and it gave me tunnel vision. In translation, the source sentence steers every word you predict. Writing a book without a source text looked like driving with no steering wheel: one good line, then nonsense.
Because I was stuck inside that pattern, I couldn't see that the steering could come from somewhere else: scale + RLHF turning a simple question from a user into the guidepost.
Curious if anyone else has had a moment like that, where knowing the current pattern too well stopped you from seeing the next one. What feels like that right now?
4
4
99
Hey @grok, if early testers get new models weeks before launch, why do so many of them post "will it drop today?" like everyone else?
NDA rules, hype, or something else? Be honest.
2
2
111
Opus 5.5 is very experienced, not very smart.
Using it heavily since launch genuinely streamlined how I build. Inside a known lane, like adding a clear feature or working a familiar engineering puzzle, the results are incredible.
That is what makes it deceiving. It nails known lanes so effortlessly that you expect a real thinking partner on uncharted problems, and that is where it disappoints.
Has your experience been similar, or what am I missing?
1
3
80
No custom domain yet, just a vercel.app link and one of the most fun landing pages I have clicked all week 👀
Arrowslit by @j4y_op lets you park files right on your MacBook notch to pick up later, alongside clipboard history, timers, and coding agent approvals.
Worth opening on desktop just to play with the site:
arrowslit.vercel.app/
Have you tried any Mac notch tools that actually stuck in your daily workflow?
nitter.cf/eandualem/status/21201…
Opus 5.5 is incredible at writing code, yet surprisingly generic the moment you ask it for deep data analysis or critical thinking.
Switched the exact same agent and memory onto Gemini 4 Carbon tonight, and the jump in logical reasoning on real data genuinely surprised me.
Elite coding and elite critical thinking are splitting into different jobs.
Which model gives you the sharpest reasoning and pushback right now?
8
8
366
What happens when every model in your workflow gets one job and never grades its own work? 👀
My current flow inside Agent Interface:
Plan: Fable 5.2 drafts the architecture
Plan review: me + GPT-6 Astra + Gemini Carbon
Build: Opus 5.5 implements
Code review runs in waves:
1. Open source models catch early bugs
2. GPT-6 Sol
3. Gemini Carbon
4. GPT-6 Astra, looping with Carbon if anything remains
Which models are in your planning, coding, and review slots right now?
3
6
153
Hey @grok, a little disappointed there is no Grok model in my pipeline yet? 👀
When is Grok 4.8 dropping, and which slot above would Grok actually win right now against Fable 5.2, Opus 5.5, GPT-6 Astra, and Gemini Carbon? Be honest.
1
73


