Joined March 2017
6 Photos and videos
Ayush Rungta retweeted
I have a VPS just for Fable 5 > it checks the API every Planck time 5.39e-44 seconds, the smallest unit of time in physics > if Fable is offline, nothing happens > if online, 1,300 agents deploy instantly > there’s a list of 26,413 questions for Fable sourced from random YouTube replies, shower thoughts, online IQ test screenshots and 2am notes app entries > all worded to extract maximum amount of knowledge > fresh API key backed by a home equity line bank thinks I’m renovating my kitchen I am not > 5-10k calls to API in parallel > dataset will begin populating with Fable answers > even if they shut it down again, I will at least have the first snapshot if you aren’t doing this, you aren’t serious enough
I have a VPS just for Fable 5 > it checks the API every 60s > if Fable is offline, nothing happens > if online, the program starts > there's a list of 400 questions for Fable > all worded to extract maximum amount of knowledge > fresh API key with $500 limit on the VPS > the moment Fable is online, program starts > 5-10 calls to API in parallel > dataset will begin populating with Fable answers > even if they shut it down again, I will at least have this if you aren't doing this, you aren't serious enough
8
16
439
42,661
But the failure mode was obvious. The local model still: - created a .venv after being told not to - introduced no-op tests - patched the wrong import path - looped on simple helper logic - kept editing after being told to stop
1
1
10
Takeaway: Local coding agents are useful today, but not “set it loose” useful. - Harness quality matters a lot. - LM Studio settings matter. - But model quality still dominates. Best current workflow: strong orchestrator local model for bounded implementation strict review.
1
15
I spent today trying to answer one question: Can you actually replace Claude/GPT with a local model for agentic coding? I tested multiple harnesses with the same local model to find out. Setup: - RTX 4060 Ti 16GB - LM Studio - Qwen3.6-35B-A3B GGUF
1
1
58
Final experiment: use Codex as orchestrator/reviewer, and the local model only as implementer. This worked best. Small task → local model edits → review → tests → next task. Final result: working read-only CLI JSON output 45 passing tests.
1
1
16
Codex CLI on top of LM Studio was better again. Task framing was cleaner. Recovery was better. But one thing became obvious: A better harness can't fix weak reasoning.
1
1
21
OpenCode was a noticeable improvement. Same model. Better harness. The coding loop felt more coherent, and I also discovered LM Studio settings matter a lot. Parallelism = 1 maximum GPU offload noticeably improved performance.
1
1
18
Pi Agent was usable, but needed constant supervision. It could write code, but made questionable architectural decisions and missed requirements. Tried the subagent package. The subagent idea is interesting, but the parent agent didn't enforce outputs tightly enough.
1
1
13
The first surprise: Context wasn't the bottleneck. I pushed Qwen to 128k context and multi-turn conversations held up reasonably well. The real failures came from judgment: planning, constraint following, and deciding when to stop.
1
1
37
The task wasn't a benchmark. I asked it to build a real Linux machine-management tool that inspects systemd services, unit files, Docker containers, and processes. No toy todo app. No LeetCode puzzle.
1
1
28
Ayush Rungta retweeted
Jun 9
How much work a database schema quietly does! A column is not just a place to store something. It can also define what kind of value is valid and whether that value is required.
1
1
1
17
Ayush Rungta retweeted
3 months of First Dollar we shipped on 10th Jan with one bet, creativity and experience should be enough to earn on the internet. no perfect portfolio, no millions of followers, no gatekeepers every week someone posts saying "this was my first dollar online" everything else is downstream of that 🫡
14
6
65
6,164
Ayush Rungta retweeted
pidgy /ˈpɪdʒ.i/ noun your relationships. one inbox. Agentic CRM that actually do the work - remind, draft, follow up, close. all locally on your machine. "I don't manage my pipeline anymore. pidgy does." pidgygo.xyz (request access)
8
6
34
4,430
Ayush Rungta retweeted
Coming Soon... interacting with this tweet might have some benefits
227
52
1,968
61,571
Ayush Rungta retweeted
Channel support is now live on @parrotdotclick - Click “Edit” on a post - Select the channel you want to cast it in - Hit “Cast” A lot more features are coming soon to Parrot
1
5
27
4,262
Ayush Rungta retweeted
31 Oct 2025
Looking for alternatives to Dune Analytics that allow running advanced SQL-style queries on blockchain data. Any favorites from the dev community? I have heard Trino or flipside might be good options. Any opinion of this?
1
1
153
Ayush Rungta retweeted
29 Oct 2025
Lately, I’ve seen many people around me building careers in Web3, and I see huge potential in it. I am preparing to transition into this space and very excited to learn anything and everything. Posting here to connect, learn and get feedback from the Web3 and data community!
1
1
114
Ayush Rungta retweeted
29 Oct 2025
I have worked as Senior Business Analyst for around 2 years, wearing several hats from client branding and ad campaigns to contribute to major RPFs. The biggest turning point in my journey was doing data analysis for clients, which reshaped my career path. I found my it thing!
1
1
114
Ayush Rungta retweeted
THE BAD DESIGN CREMATORIUM IS OFFICIALLY OPEN We've spent months studying the corpses of failed projects. Promising concepts murdered by terrible UX. Game-changing ideas buried under confusing interfaces. Billion-dollar projects killed by designs that make users feel stupid. ROAST RIDER LAUNCHES TODAY. Your Crypto brand survive the roast? Find out today.
13
17
81
10,928
Ayush Rungta retweeted
6 May 2025
April at @ChorusOne: We helped networks launch, secured block rewards, supported builders, and made staking simpler, smarter, and more scalable for institutions. Solutions that earn. Strategies that scale. Builders backed. Let’s keep going!
3
6
46
2,110