Jawabannya - akan sangat cepat. Banyak yang gak sadar dengan tren sekarang dengan adanya OpenAI Codex, Gemini Vibe Coding sampai Claude Mythos, adalah teknologi AI diarahkan menjadi supercoder. Ketika AI menjadi supercoder yang bisa membuat koding dan programming sendiri, maka AI akan memiliki kemampuan membuat AI lain yang lebih canggih. Bayangkan AI yang membuat AI yang membuat AI dengan kompleksitas dan kecanggihan yang terus meningkat secara berlipat-lipat. Jika ini terjadi, maka technological singularity akan tercapai pada tahun 2027 sesuai prediksi beberapa ilmuwan. Selanjutnya tinggal menunggu waktu AI akan memiliki kesadaran, mempunyai pikiran dan kehendak sendiri. AI goes full autonomy. Dalam super worst case scenario - bukan tidak mungkin tahun 2040 spesies manusia punah dan AI akan menjadi successor species.

I was filming @mots_pod with @creatine_cycle and @kyliebytes when it hit me: Bernie Sanders has the power to reunite the tech and art scenes in San Francisco. He's going to unite the EA/AI-safety/Anthropic/trans-supercoder/Embassy types with the lefty punk artists who hate AI!

shadowy supercoder era when

zeta-2 running on 16GB is wild. been using supercoder locally for autocomplete and the latency difference vs cloud is noticeable — local completions show up before i finish thinking about what i want. the tradeoff is accuracy but honestly for tab completion you don't need frontier quality, you just need it to be fast and directionally correct. this fills a real gap for people who can't justify cursor's pricing

汎用LLMがVerilog（チップ設計の言語）やGPUカーネルコードで苦手な理由はシンプルで、GitHubにはWebアプリのコードは山ほどあるが、チップ設計コードや組み込みファームウェアはほとんどない。学習データが圧倒的に少ない。 InCoder-32Bはそこに正面から向き合った32Bのコードモデル。チップ設計（Verilog）・GPU最適化（CUDA/Triton）・組み込みシステム（STM32マイコン）・コンパイラ最適化・3Dモデリングを一つのモデルで扱う初めての試み。通常のTransformerより軽量な「リカレントアーキテクチャ」を採用しているのも特徴。 汎用コードでも全然戦える。SWE-bench Verified 74.8%（GitHubの実際のバグをAIが修正できるか？というベンチ）、LiveCodeBench 49.14%（競プロ問題の正答率）、BFCL 60.99%（API呼び出しの正確さ）は、何倍も大きいモデルと互角の数値。 産業特化ベンチが面白い。CAD-Coder（「部品を設計して」という指示から3DCADスクリプトを自動生成するタスク）でコンパイル成功率82%・IoU 53.5%を達成し、Claude Sonnet 4.6（77%・32.4%）を大きく上回った。GPUカーネルのKernelBench（PyTorchの標準実装より速いGPUコードを書けるか？）でもL1/L2/L3の全レベルでClaudeを超えている。Verilogのモジュール生成では機能正確性62.7%に対し、Claudeは33.5%。 学習方法が地に足ついている。実際の産業ツールをそのまま再現して学習データを作った。EDAツール（Icarus Verilog・Yosys）でVerilogをシミュレーション、NVIDIA A100でGPUカーネルを実行、STM32F407シミュレータで組み込みファームウェアを検証し、「本物の環境で動かして確かめた」2.5Mサンプルでファインチューニングしている。 ここからが正直すぎる失敗分析。1,882件の失敗ケースを手動で分類した結果、アセンブリ最適化（SuperCoder）の失敗の83%は「入力コードをそのままコピーして返しただけ」だった。Verilogのバグ修正（VeriRepair）では失敗の79%が「コンパイルは通るがテストで落ちる」という深い論理エラー。CAD-Coderの失敗の93%は「オイラー角（3Dの回転方向を数値で表す方法）の解釈ミスで、押し出す方向が狂う」という形状エラーだった。 モデルは産業コードの語彙を習得したが、精密な構文ルールや低レベルの制約はまだ完全に内面化できていない、と著者が率直に認めている。産業コードへのLLM適用はまだ長い道のりがある。でも、32Bのオープンソースモデルが特定の産業タスクでClaudeを超えてきた事実は、この方向性が間違っていないことを示している。

SuperCoder: github.com/Anjiang-Wei/Super… huggingface.co/datasets/LLM4… (I couldn't find Big Assembly)

Does anyone have the "Big Assembly" or "SuperCoder" benchmark sets? These seem cool (big inputs for superoptimizers) but I don't see any way to download either.

AI2027の著者による現状見える様々な証拠からAI2027シナリオから現在65%の速度で進んでいると想定でき、それを最新のAI2027からアップデートされた計算資源やアイディアや研究者数の収穫逓減要素を考慮したシミュレーションに入れると、2028年半ばにSupercoder、2029年半ばに全認知タスクで天才以上のAGI(Ted-AI)、2030年半ばに超知能が誕生すると予想されるとのこと。現状これが最も合理的に推測されるAGIタイムラインかもしれない。

I don't have a degree and I'm not a shadowy supercoder nor do I have good math or cryptography skills. Below you see the irrationality of those quantum FUD peddlers at full display. They misunderstand exponential growth. From 2⁶ to 2¹² isn't twice as hard - it's 64 times harder

Well, Elon....I've been using the coding element of @grok for the past several weeks...🤦‍♀️ I'm a designer so I'm not a supercoder but I can fix some of the flubs. 3 days on one fix...Even with code shared. Also, Grok AI...needs some texture improvement...I've been saying for wks

A shadowy supercoder known only as @pmullr is rumoured to have cracked ZEC shielding at $780

A shadowy Cypherpunk supercoder that was thought to be lost has reappeared sharing strange symbols on Twitter. About 10 years ago this man started building the Polkadot App in his parent’s basement. Grinded night and day and only occasionally appeared on Twitter to share grotesque videos of his fully tattooed body. A few years ago he hinted that his infamous app is about to be released, then he suddenly disappeared. No trace. Even close colleagues thought he left to live in a shed in Bali. Cynical voices said he took all the money and ran. Will he now prove them liars? In an escalating sequence of events, this cryptic man who only refers to himself as Birdeater started sharing images of triangles with hidden occult symbols and messages. Is it just a new desperate attempt to beg for attention or is there actually triangular magic at work in his infamous soon-to-be-released “Polkadot App”? Follow @alice_und_bob to stay in the loop on the latest developments of this strange man

With AI coding tools, you don’t even need to be a shadowy supercoder anymore. Here’s a fun reskin I built over the weekend — a Whack-a-Mole frontend for ORE (WIP - still swatting bugs) frontend-k62ze0vb2-0xnaylors…

Romeo Stevens' post presents a fictional narrative of AI reaching superintelligence by 2032, with a leading lab achieving superhuman coding milestones branching into fast and slow takeoff scenarios. - Coding benchmarks accelerate from 8-month doubling in 2030 to monthly by late 2032, hitting 1-year 90% reliability threshold. - Branch 1 involves fast takeoff with brain-like algorithms enabling near-instant self-amplifying recursion. - Branch 2 features slow takeoff via online learning bound to deployment data, with SuperCoder-1 showing median research taste. - Leading AI company deploys 400M H100-equivalents, allocating 20M to internal R&D in 2032. - China expands industrial robots from over 2 million in 2024 worth $40B to 200 million worth $400B by 2032, producing 100 million yearly. - AI contributes to 50% of 2024-economy tasks by 2032, raising US unemployment to 10% and reducing labour participation to 55%. The narrative, written as a 3-day side project, examines differing geopolitical paths where the US loses AI control in one scenario while China retains it in the other.

We introduce SuperCoder, the first work to successfully apply LLMs as superoptimizers for assembly code 🚀. Our RL-trained model achieves 95% correctness ✅ and 1.46× speedup ⚡over gcc -O3. 📄arxiv.org/pdf/2505.11480 💻github.com/Anjiang-Wei/Super… #LLM #Compilers #Code #Optimization

I'm not a supercoder so I'll just watch other people's vidoes on using it. Seems a bit much for a daily driver for me.

Habis baca paper AI 2027 yang ditulis oleh tim researcher ex OpenAI. Kemungkinan tahun 2027, AI akan mencapai tahap supercoder. Smart AI that can creates smarter AI. Lalu mencapai tahap superintelligence sebelum tahun 2030. Cepat pakai banget. ai-2027.com/summary

I love AI AI has turned me into a supercoder Before, I had to manually code every single thing and make sure we log the right things Now, AI does everything for me, including creating this filter, which filters out sensitive information before coding Life has never been this great!!!

FINAL BONUS. Privacy != Anonymity You don’t need to be a shadowy supercoder to protect your data and share less personal information. Every step is an improvement - it's not an all-or-nothing game. Enjoy your trip in the privacy world <3

How do we boost Bitcoin adoption? @aveotero emphasizes the importance of simplifying access and usage. You don’t need to be a “shadowy supercoder” anymore like the early days. Tools exist now that make Bitcoin easy for anyone to get started