Digital Nomad I’m the Flash ⚡️ but for the frontier
Joined August 2022
- Tweets61
- Following158
- Followers27
- Likes312
Abhinay Reddy retweeted
I'm so sorry, @DuoLingo. They copied us too #Solidarity
Abhinay Reddy retweeted
its so funny how accurate and inaccurate it is at the same time 😭
why does it throw so much unnecessary text literally everywhere
Abhinay Reddy retweeted
AI is much bigger than most people realize.
We currently have multiple $1+ trillion companies switching spots as the "best" AI model on a monthly basis.
OpenAI just went from "falling behind" in the AI race to reaching Artificial General Intelligence (AGI) in a matter of weeks.
Meanwhile, frontier models from OpenAI, Anthropic, Google, and xAI continue to leapfrog one another across major benchmarks and capability categories.
When you have multi-trillion dollar companies losing and regaining the lead on a monthly basis, that's not normal.
That's indicative of a historic technological revolution, with hundreds of billions of dollars in annual CapEx spending and trillions of dollars in AI infrastructure investment planned over the coming years.
We are in the midst of a generational shift in technology.
Take advantage of it.
That means Gemini can simultaneously be:
Excellent autonomous coding agent
+
annoying interactive coding model
Also, DeepSWE measures model + mini-SWE-agent harness + reasoning setting, not simply raw model intelligence.
So I wouldn't say the benchmark is wrong.
It means Gemini has very high problem-solving throughput.
But benchmark equality does not mean equal engineering judgment.
so a model can scores 74%, but still needs you to babysit it much more.
And once a developer loses trust, because the model rewrites too much, misses intent, gives inconsistent fixes, or takes strange paths, then the benchmark numbers means nothing..
Claude and codex developed a reputation for codebase understanding + clean edits + strong reasoning + interactive collaboration.
Gemini’s problem is not necessarily intelligence.
It is that capability has not always translated into developer trust per interaction, because Developers care whether they would actually let it drive.
Abhinay Reddy retweeted
I challenge this a bit
(a) Codex was started as an internal tool. The goal was explicitly to build what was the highest leverage tool to accelerate infrastructure buildout.
(b) I don’t think the Meta vs non-Meta is particularly enlightening here. Most of us who came from older tech companies were surprised at how far OpenAI had taken Python scripts and had a desire to scale OpenAI further.
(c) A ton of infrastructure and internal tools were built over the last few years, it’s not the case that we were just waiting for AGI to appear.
Abhinay Reddy retweeted
"man GPT-6 Astra CRUSHES fable. anthropic IPO is screwed"
"yeah benchmarks are crazy, AGI is here"
"have you tried it?"
"no i don't have access. you?"
"no."
