Pinned Tweet
In the last few years, I've been reading books and meticulously taking notes on them.
I'm sharing my notes for free.
If you want to learn something useful, head on over:
bagerbach.com/books
Some thoughts on Gemini 4 Carbon:
Pros:
- Best at writing, anecdotally and in my evals. Still not great.
- Grasps intent better than even Claude models (vs. Opus 5.5, Fable 5.2)
- Occasionally outperforms Opus 5.5 in quality of solution
- Usable for review. Can suggest designs and not overthink, unlike GPT models
Cons:
- Very, very slow overall
- High rate of hallucination compared to literally any other frontier model
Used on High reasoning. Used the model in both T3 Code with Antigravity via ACP and in Amp.
@_mohansolo @OfficialLoganK i want gemini to flourish, hope this feedback is useful 🙏
2
1,026
It's clearly the best Gemini model so far at coding. But is that enough?
With the amount of hallucinations and how slow it is, I'm not so sure.
Let's take writing docs as an example.
+ best model at writing
- slow and has issues getting the assets (driving the app to take screenshots, gathering user flows, etc.)
- hallucinates often, so information has to be double-checked
So to actually use it for writing docs:
- another model finds all the requisite material and take screenshots etc.
- gemini writes docs
- another model reviews for factually incorrect information or other hallucinations
That's not a particularly nice workflow.
1
10
1,947
This puts it in a very weird spot.
If you could only get one subscription, it wouldn't be my first choice.
But if you're using multiple subscriptions for development work, their ultra subscription is a good deal.
Not only do you get Gemini, you also get YT premium, a bunch of storage, etc.
1
6
1,439
Introducing Claude Fable 5.2 and Claude Mythos 5.2, designed to tackle the most daunting technical challenges.
They’re capable collaborators on ambitious projects in many fields, and show improvements in areas like coding, reasoning, and vision.
Christian Bager Bach Houmann retweeted
Replying to @brian_lovin
Skill issue.
New side projects: change legislation, robotics, hardware, new kernels, new data stores, rockets/drones, biology…
Old side projects: notes app, read it later app, marginally better coding tool, weather app, todo list app, camera app, hacker news app, voice dictation app…
Great article about some of the problems we get to solve these days. recommended read.
software engineers are going to have to tackle a new set of problems to unlock powerful multiagent systems to solve the world's hardest problems - right now is one of the most interesting moments in history to be an engineer.
2
266
QuickAdd for Obsidian now has an official Raycast extension.
Every workflow you've already built works.
Now it's faster and easier than ever to capture or create notes!
3
12
764
Christian Bager Bach Houmann retweeted
It's mostly me writing a great prompt that explains what I want and a direction.
The idea is to get out of the loop outside critical moments, which is usually when defining what needs to get done, and before shipping it to users.
I.e. you need to set the direction, and you need to understand what's being shipped and agree that it's a good change.
To get further out of the loop, you need to trust the model outputs.
So t3-pstack runs judges, validation & verification, and a ton of other stuff automatically to ensure a high quality & that things work. I work on pretty critical systems so I spend a lot of time on ensuring everything is secure, performant, and has a high quality bar.
Many of the skills are invokable by the model, and e.g. opus 5.5 is great at using them. So it's not like you'll be using slash commands all the time.
Most of the 'manual' work is in creating great environments. E.g. making it possible and easy for the agent to check its work, get great feedback, run uninterrupted, etc.
Your ability to design great environments for agents will make the returns on using skills like t3-pstack compound and you'll ship like never before. At least, it has worked for me - even before pstack.
I ported @poteto's pstack to @t3dotcodes.
Feels absolutely incredible to use.
It's native T3, meaning we use the T3 Code built-in tools and tune the stack entirely to t3.
I'm trying to keep it faithful to pstack while optimizing it for t3.
github.com/chhoumann/t3-psta…
Worked on getting sandboxes into T3 Code
Instead of running on main or a new worktree, it spins up a fresh boat.dev VM, run the agent there, and adds it as a remote environment on all devices
No more managing memory on my machines or balancing worktrees
Works by having your T3 server hold a Boat key and owning the lifecycle.
Creates the VM, installs T3 inside it, and every client you've paired picks the sandbox up automatically.
It's just another T3 environment.
Still lots and lots of work to make this good, but it's a cool prototype
2
30
2,328
@theo would love to have this upstream!
My machines keep oom killing stuff because I have too many agents working at the same time. T3 is helping me spin up more ideas than I have compute for
8
46
1,897
Christian Bager Bach Houmann retweeted
"The invention of the boat reduced the importance of swimming, but enabled the discovery of new lands."
Beautiful analogy from @ylecun about the opportunities and changes with AI, in mathematics and in science more broadly.
This aged very poorly
Grok Bot is a great product and I’m finding it very useful.
Massive respect for the decision to use the best model for the job, not just their models
I really like Grok Bot, but the worst part of it is that it uses Grok. The models frequently underperform or misunderstand, so I end up not trusting it to actually work for me.
I can’t wait for OpenAI’s solution here.
1
269














