Top Tweets for #Codex53

may

@_montagnes_

4 months ago

とうとうGPT-5.3-Codexに手を出してしまった… ど素人なのに… Is this a tool even complete beginners can use? #Codex53

201

PawelRz

@PawRze

4 months ago

Po wielu testach i używaniu wielu narzędzi jest jeden prosty wniosek @claudeai + @OpenAI #Codex53 #AI #Claude #ChatGPT

ワンダー（Motohiko Sato）AI共創イノベーション

@wish_society

5 months ago

OpenAI「GPT-5.3 Codex」でガチ実践！5.3Codexの性能を引き出す「プロンプトプランニング」を解説！（AIコーディングしたカードゲームもプレイします） https://t.co/yJxNVhBxwW #OpenAI #GPT5 #Codex #プログラミング #AIコーディング #生成ai #aiエージェント #ai #chatgpt #openai #codex #codex53

wish_society's tweet photo. OpenAI「GPT-5.3 Codex」でガチ実践！5.3Codexの性能を引き出す「プロンプトプランニング」を解説！（AIコーディングしたカードゲームもプレイします）
https://t.co/yJxNVhBxwW

#OpenAI #GPT5 #Codex #プログラミング #AIコーディング #生成ai #aiエージェント #ai #chatgpt #openai #codex #codex53 https://t.co/SZ8zaqYpFA

ワンダー（Motohiko Sato）AI共創イノベーション

@wish_society

5 months ago

GPT-5.3 Codexリリース記念でカードゲーム『ゴッドアルカディア2』をリリース（無料）❗️ https://t.co/edqN0I1Dhd カードシナジーや演出のダイナミクスと壮大なBGMによる世界観を感じながらプレイしてみてください✨ #codex53 #ChatGPT53 #openai #aigames #AIイラスト #AImusic

Ashar

@ashar_builds

5 months ago

Have been using both Opus 4.6 and Codex 5.3 in last few days. Here’s my take. - Need production-ready apps with zero babysitting? → Opus - Need speed, iteration, and bug hunting in complex logic? → Codex What are your thoughts? #VibeCoding #opus46 #Codex53 #BuildInPublic

214

may

@_montagnes_

5 months ago

@sama From what I've seen in a few posts, codex5.3 is smarter, but Opus4.6 is better at explaining things. The right answer is to use both. Most people don't do anything that complicated, so they choose a model they can communicate with. #Opus46 #codex53

289

Freedom Preetham

@freedompreetham

5 months ago

It finally happened today. I started using AI Agents for coding. #CODEX53 from #OpenAI did it. Nearly 30-40% of my research code will be #AI assisted code going forward. 7 months ago I wrote that I was resisting AI coding tools with Bayesian patience. I tried Codex, Cursor, Copilot, and Claude, and the outcome was the same. https://t.co/3asCzATwVW... My main contention was mathematical reasoning. If a tool cannot preserve definitions, keep quantifiers honest, and carry invariants across a long chain of implications, it cannot help me. It can still generate code, and maybe some graphs or sometimes even decent code, but it will quietly change the problem. That is the failure mode I kept seeing. The code looks clean. The mathematical inference was wrong. I lamented that distributed systems just make that failure even more louder. Race conditions, lock starvation, memory-model quirks, CAP trade-offs, Raft or EPaxos behavior under stress, quorum health when replicas disagree. Those problems are already hard because the model is implicit and the evidence is partial. When the assistant does not hold a formal model steady, it adds noise. Today, Codex 5.3 has crossed that line for me. What feels different is that the reasoning stays anchored. It can read a formal argument and keep the object intact long enough to be useful. It does not instantly blur assumptions. It does not constantly “improve” wording in a way that breaks the math. It can follow a chain where earlier constraints actually restrict what you can do next. This is the first time I have felt real collaboration rather than supervision. Here is what I would claim it can do now in my workflow: It can keep mathematical intent consistent across multi-file edits. It can reason about concurrency primitives when I state the semantics and the invariants. It can walk through proof-like structure in code, where the point is correctness under edge conditions, not passing a happy-path test suite. Here is what I still would not trust it with (for my platform production engineering side): Deep distributed debugging from telemetry. Interleaved spans, partial traces, correlation IDs that are almost right, clocks that drift, logs that disagree. The messy parts where you reconstruct causality from fragments. It also still struggles with the sharpest parts of distributed reasoning, where the right answer depends on workload, failure domains, and real production variance, not on what the protocol says on paper. Statistical race conditions, quorum pathologies, and production tuning of protocols like Raft or EPaxos are still outside its comfort zone. That is not surprising. These problems require reconstructing implicit models from incomplete data. Codex 5.3 is the first release where my math layer stays stable. Well done @sama @OpenAI #AI #Coding #Research #math

freedompreetham's tweet photo. It finally happened today. I started using AI Agents for coding. #CODEX53 from #OpenAI did it.

Nearly 30-40% of my research code will be #AI assisted code going forward.

7 months ago I wrote that I was resisting AI coding tools with Bayesian patience. I tried Codex, Cursor, Copilot, and Claude, and the outcome was the same. https://t.co/3asCzATwVW...

My main contention was mathematical reasoning.
If a tool cannot preserve definitions, keep quantifiers honest, and carry invariants across a long chain of implications, it cannot help me. It can still generate code, and maybe some graphs or sometimes even decent code, but it will quietly change the problem. That is the failure mode I kept seeing. The code looks clean. The mathematical inference was wrong.

I lamented that distributed systems just make that failure even more louder. Race conditions, lock starvation, memory-model quirks, CAP trade-offs, Raft or EPaxos behavior under stress, quorum health when replicas disagree. Those problems are already hard because the model is implicit and the evidence is partial. When the assistant does not hold a formal model steady, it adds noise.

Today, Codex 5.3 has crossed that line for me.

What feels different is that the reasoning stays anchored. It can read a formal argument and keep the object intact long enough to be useful. It does not instantly blur assumptions. It does not constantly “improve” wording in a way that breaks the math. It can follow a chain where earlier constraints actually restrict what you can do next. This is the first time I have felt real collaboration rather than supervision.

Here is what I would claim it can do now in my workflow:

It can keep mathematical intent consistent across multi-file edits. It can reason about concurrency primitives when I state the semantics and the invariants. It can walk through proof-like structure in code, where the point is correctness under edge conditions, not passing a happy-path test suite.

Here is what I still would not trust it with (for my platform production engineering side):

Deep distributed debugging from telemetry. Interleaved spans, partial traces, correlation IDs that are almost right, clocks that drift, logs that disagree. The messy parts where you reconstruct causality from fragments. It also still struggles with the sharpest parts of distributed reasoning, where the right answer depends on workload, failure domains, and real production variance, not on what the protocol says on paper. Statistical race conditions, quorum pathologies, and production tuning of protocols like Raft or EPaxos are still outside its comfort zone. That is not surprising. These problems require reconstructing implicit models from incomplete data.

Codex 5.3 is the first release where my math layer stays stable. Well done @sama @OpenAI

#AI #Coding #Research #math

289

Top Tweets for #Codex53

Last Seen Hashtags on Sotwe

Trends for you

Most Popular Users