AI agents engineer | Shipping games, apps and agent experiments in public
Joined March 2024
- Tweets6.5K
- Following399
- Followers772
- Likes2.9K
Pinned Tweet
I build games and apps with AI agents and post what actually happens. A few things I did recently:
1. Got glm in @cline to build a racing game where every texture and sound is generated, nothing downloaded.
2. I simulated an entire company to see where agents break.
3. Made a cel-shaded minecraft clone for my sister's birthday. One prompt, 3M+ tokens, ran for 4 hours.
4. And my agent saved my pc from a bsod that would have cost me 147$ for 😁
All of them are in replies 👇
more in hilights.
4. BSOD Error fix
My agent helped me recover my faluty pc for free that would otherwise had costed me 147$. A real story!
Today I was watching my college lectures when suddenly my screen flashed and my pc went into a bsod ( blue screen of death ) error.
Though I was aware of how to recover it , I failed to do so , as everytime I try to open the pc, it would ask for a recovery, which I never wanted.
Luckily during one iterartion , I stepped into the actual error code, took that screenshot and fed into my agent I use on my mobile.
Also I added that I don't want to loose any data , and don't want to go for recovery , is there a way to do it.
My agent went through its internal monologue and came back saying yes there is 30-70% chance of success- failure.
As I don't had anything to loose , I took up that chance and it instrcuted me the instuctions:
> Head to the advance startup
> Open command prompt
> Type sfc/scannow & wait
> Type chdsk /f/r
> Wait till complete
> Type exit
> Click Continue (Exit & continue to Windows 11)
I ran commands (I knew what they do) , waited and after 2 hours , I was able to login and none of my file where deleted , neither I had to reset my pc.
So out of curiosity , I asked my agent what likely happened and learnt that:
> my pc's one of the driver tried to access the memory , it didn't had access to, which curropted the system files and forced a device into bsod state.
I dug deeper and found that the real culprit for this was windows "Core Isolation" setting being turned out.
Although this is a safety features to restrict memory corruption but sometimes this causes issues for drivers.
Furthermore the 2 command I ran, were not normal commands:
> /scanfnow - system file checker , checked for all the protected windows os files and replaced the damaged, missing or corrupted one with clean cached copies
> /chkdsk/f/r - scanned storage drivers , to verify file system intergrity, detect, fix and recover (/f/r) storage errors, while marking bad clusters as unusable , so no further writing on it. Done in 5 stages.
In short , a well though , planned and structured recovery setup that feels like magic , but it all works on maths and CS fundamentals.
That's the basic overview of what all happened , and how agent fixed it all.
Bottom line is that:
AI isn't going anywhere , if you fear from it , your job might be.
But if you adapt it in your workflow and use it to gain knowledge about what it does, how it does it and why it does it , you will become 10x +.
Thanks for reading, hope this even act as a solution to people facing the same issue.
More 👇
I recently setup my android app developer factory , the mazor hurdle was to use it in a headless environment with minimal ui usage ( no android studio).
Android studio became the answer, and I integrated it my workflow, helps a lot with building apps automatically.
You enter the requirements , factory builds a specsheet, requirements , setup a environment (android cli + kotlin) , build the app using tdd approach , launch emulator test itself and then deliver the final product with a doc explaining the build.
I think android device hub can be integrated in a similar way, for letting multiple agents work in parallel on different apps and test them realtime , without any dependency and rance conditions.
Introducing the Android Device Hub, a streamlined app to view and interact with all your local and remote devices and emulators, without needing to open an IDE.
We launched an early version of it last week at the @AndroidDev Summit, and would love your feedback on shaping its future, as we add familiar features from Android Studio, and also explore some new ones.
Try it out with `android device hub` via Android CLI!
This one is actually useful for evals , half the time I am not sure which model the api really served me after a reroute.
If I can verify the response came from the model I paid for , my eval numbers finally mean something.
New Anthropic research: progress on provable inference for large language models.
Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.
These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.
For our full write-up, see here: anthropic.com/research/prova…
Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: github.com/anthropics/FlashP…
We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (rdalal.github.io) and Ben Mayer (@transmissions11).
Funny how the 5-hour limit was quietly the best eval I had 😁
Every time I hit it, I had to stop and check what the agent actually did in those hours.
Now Claude Ultra drops the 5-hour limit for verified accounts, so nothing forces that check anymore.
> no wall means no pause
> no pause means nobody reads the logs
> and that's how your weekly cap is gone by midweek
The model was rarely the problem, the missing checkpoint was.
Will you add your own stop points, or just let it run?
Context caching lets you store frequently used input tokens in a dedicated cache and reference them for subsequent requests , eliminating the need to repeatedly pass the same set of tokens to a model.
Been learning from this a few month back and the curriculum is hands down the best for practical developement.
Repo is constantly updated too with new and important stuff required along with new forms of learning material!
Thanks to developer who made it 😁
Tier 1 city airports and flights are the best place to network and have meaningul conversations.
Met a guy today , who was suppose to miss his flight , due to late flight arrival from Zurich. Glad he didn't.
During our conversation he shared , he is senior head detection enginner at sentiel one and take interviews. I got curious.
We had a lot of convo , but here are main pointers ( in no particular order)
> Never say that it's company internals , can't share , it reduces the confidence of the recruiter in you.
> Share what you have worked on within organisation. In case you haven't worked on a tool , never say no , honestly mention what you know about that tool. Raises credibility .
> Never keep on applying like blindly. , if you are getting 5-10 rejections ,it's time to think, optimise the skills and then reapply.
> Rejections are real , embrace it , learn from failure and improve upon it , don't take it as setback.
> LinkedIn very good place for cold dms , reddit and X for marketing . Utilise it.
> Sales people have a lot of penetration in the organisation, they know details of internal meeting (who will be kept and removed). Connect with them (much better chance then blindly shooting resumes)
> If you are in 1/2nd year explore as much as you can , and find you passion and get drowned in it.
> Focus on fundamentals: ai can do things , but you need to know fundamentals to debug.
> Juniors goals going to disappear, it's inevitable. High agency folks stays.
> Focus on domains like cybersecurity , data scientist , hardware , low level programing (kernel, drivers, sockets, db, etc) , quantum computing and biotechnology.
> Always keep the raise rate as per market standards, builds credibility (hr will try to keep that low , don't accept)
> We also taked about stocks , esops , and how to use them beneficially to build an forturne.
> And finally how he is building a product and can use some helping hand ( yup may be my next potential client 😁)
Most of you might already know these, but wanted to share my experience with X's amazing community.
For a fun fact : the guy is not available on any of the social media , he says it creates work distraction.
Peace ✌️
Folks always recheck you banking details before adding @Razorpay integration.
Spend 2 days going around why users are not able to pay to the bank account, only to discover the IFSC check was failing and surprisingly razorpay never complains for wrong IFSC code.
Did one line change and it's now fixed!
harsh retweeted
SpaceXAI engineer Lauren Tan:
"99% of people are using GrokBot with less than 1% of its real power
They run a single agent without loops or graphs"
"I'm running a fully autonomous team of 20+ GrokBot agents
One Chief of Staff, one PM, and 20+ workers
That's the new engineering stack"
GrokBot → Chief of Staff → PM → Workers → Autonomous Fleet
In this 1-hour session, a SpaceXAI engineer shows how to build a complete AI agent team from scratch
This workshop is worth more than most $1,000 agent engineering courses
Bookmark it and watch today
Then read the full article below
