Matt Ambrogi
@matt_ambrogi
Followers
788
Following
12K
Media
468
Statuses
3K
Programming, product, and machine learning.
Joined November 2019
"The doers are the thinkers" is more true than ever with AI. If you're working somewhere where the AI strategy is set by people who haven't ever built an AI feature themselves, it's not going to go well. Moreover I'm skeptical of any AI strategy except move fast and try a lot.
0
0
3
People are mad that {insert model} can't one shot a poorly scoped feature for them. But a year ago being able to make a few edits across multiple files was state of the art.
0
0
0
"Two kayakers paddling on a glowing lake under a breathtaking Milky Way sky, framed by dark forested mountains.". Create images and videos in seconds with Grok Imagine.
95
46
724
Don't forget we've come a long way in one year.
1
0
2
This is simultaneously a huge problem and a huge help for what I'm working on. Crazy latency on simple tasks. But extremely comprehensive context and tool utilization. Hoping strong prompting and exploration of the new api params can overcome this.
I'm noticing that due to (I think?) a lot of benchmarkmaxxing on long horizon tasks, LLMs are becoming a little too agentic by default, a little beyond my average use case. For example in coding, the models now tend to reason for a fairly long time, they have an inclination to.
0
0
1
First round interview should literally be "tell me how you would build {insert AI feature}".
0
0
0
If I only got one 60 minute interview with an AI engineer before making a hiring decision I would much rather spend it talking to them than doing any live coding problems.
1
0
3
Oh my gpt-5 just made an agent I've been working on 1000% times better. Amazing how good it is at instruction following and tool use. Haven't seen such a big jump in what you could build on top of these models since 3.5 -> 4.
0
0
1
Underrated gpt-5 impact:. Just plugging in this model seems to immediately solve multiple significant issues in an agentic feature I’m working on. The downstream impact to consumers if that’s the case for developers working on thousands of products is huge.
0
0
1
Pausing all this “ai engineering” until I see if gpt-5 fixes everything for me
0
0
2
RT @yaroshipilov: AI makes it much easier to learn a subject deeply fast. Went from absolute zero to matching with the nation’s leading a….
0
5
0
Ok forgot Notion, this is A+.
New: Notion AI can now actually use Notion. It handles multi-step tasks like editing multiple pages or updating entire database—like a power user, built in. Rolling out over the next few weeks.
0
0
0
Why can’t I create a copy / fork of an existing chat thread from a specific message in chatgpt or claude. Want it to open in a whole new chat thread.
0
0
0
I’m weirdly interested in product / engineering strategy for building AI into incumbent b2b saas products. AI is a sustaining technology for these companies. Executing it well has created an opportunity to blow past competitors like never before.
0
0
4
And often it's almost impossible to even say what the right thing for the agent to do is. Early on you can't even form proper eval sets. You need to have data you understand well and to just test a bunch and go depth first debugging when you notice something weird.
0
0
0
Weird thing about building agents: . Getting it working is maybe 10% of the battle. The other 90% is making it work well across a basically infinite space of things users might ask it to do.
1
0
1
And notably, if I hadn't asked the first question and had asked to just jump right into the summary, it would also pick a different set of files. How do we know which is the right set of files?.
0
0
0
6. Example of the bizarre things you run into building an agent:. Changing the use case for sake of explanation. Imagine agent has access to a small code base with 20 files. Ask the agent: if you had to create a summary of this codebase and could only consult 5 files, which.
1
0
0
One of the most important things you need to be a good AI Engineer is something I’d call model intuition. Especially building agents. You need to feel “oh the model is gonna fall short here because of X and I can test that by doing Y and fix with Z.”.
0
0
2