Tymon T

Building bespoke agentic workflow automations @ xCTO Batmaid

[Tymon T @tymonPartyLate]

Thanks for the invite @ClaudeDevs
Honored to have sneakily made it into the thumbnail 😅

I'm also blown away that "ultracode with dynamic workflows" is the standard way to code now. Everyone knows it. Everyone is a loop manager now. Things move fast.

[ClaudeDevs @ClaudeDevs]

Thanks for building with us at Claude Build Day!
300 builders spent their Saturday with us in SF, and we're blown away by what you made in a single day.

Here are all the winning demos: 🧵

[Tymon T @tymonPartyLate]

@conductor_build Congrats! Pls share whats on the roadmap. More orchestration, less tab switching would be great :)

[Will Manidis @WillManidis]

https://t.co/bDCdltFo0A

[Tymon T @tymonPartyLate]

@bryan_johnson No advice. But I do hope you’ll publish a detailed trip report afterwards

[Tymon T @tymonPartyLate]

When GPT 3.5 came out everyone was like: this is system 1 thinking, we need to figure out system 2 thinking.
Now GPT 5 is out and completely lacks awareness of its own thinking. Agentic AI doesn't work except for coding. What happened to building system 2?

[Tymon T @tymonPartyLate]

@steipete that debugging process must have been hell. Can you even attach and use breakpoints?

[Tymon T @tymonPartyLate]

@nearcyan @kevinroose "All of humanity's problems stem from man's inability to sit quietly in a room alone"

The world would be so much better if they just learned how to meditate

[Tymon T @tymonPartyLate]

@sama step 1: create chatgpt
step 2: chatgpt accelerates death of the internet by powering bots and spam
step 3: chatgpt is the new internet

[Tymon T @tymonPartyLate]

@iannuttall latency is also an important factor, make the tokens flow faster from Opus! :)

[Tymon T @tymonPartyLate]

@iannuttall After a brief flirt with Grok Code, went back to Opus too. Missed the ability to actually get things done

[Tymon T @tymonPartyLate]

@ericzakariasson Went from having 2 projects with 2 agents each, back to a single agent.
The flow is best if you can mind-meld with the machine, interrupt it when it makes a wrong turn and have a clear mental map of the diffs.
Otherwise you spend all your time re-reviewing and fixing slop.

[Tymon T @tymonPartyLate]

When the economy must grow perpetually on a finite planet, it eventually has no choice but to extract more value from human life itself - colonizing our time, attention, relationships, and freedoms as the new frontiers for profit when natural resources and traditional markets reach their limits.

[Tymon T @tymonPartyLate]

@badlogicgames there is zero brand loyalty in the AI race. Switching cost is zero. People will use whatever the best new thing is

[Mario Zechner @badlogicgames]

Interesting.

[Tymon T @tymonPartyLate]

is it a conspiracy though? It is confirmed they downloaded pirated books from libgen and trained on them.
Model performance is all about the data and back when claude 3.5 was becoming great for coding, they didn't have that much synthetic data to train on.
Its not far fetched to think they obtained a great code dataset unavailable to other labs

[Tymon T @tymonPartyLate]

@SeekingAnon Memories are still an object of the material world and contained within the physical body.
Even if your consciousness can travel, it will do so without your personality and memories.

[Tymon T @tymonPartyLate]

Switching from claude code to grok code fast in Cursor. The speed makes a huge difference, even if it is a bit dumber.

[Tymon T @tymonPartyLate]

@cloneofsimo Big fan of cc. But there is a growing feeling that it must be something shady. Like training on stolen code, justified by the glorious future this will enable. Training on libgen is apparently ok so why not go deeper.

[Tymon T @tymonPartyLate]

@gregpr07 Next up: chrome was too big of an abstraction so we built our own html parser, js runtime, … 😄 good luck with your epic efforts!

[Tymon T @tymonPartyLate]

@jakozaur @GergelyOrosz Cursor for X is no longer cool. You have to do Claude Code for X now 😅