arXiv paper: a novel task linking small text snippets to the exact code segments they describe - and vice versa.
Dataset includes auto-generated GPT-4 descriptions for training plus manually annotated out-of-domain test sets. The gap between in-domain and OOD performance tells you something important about how well these models actually understand code.
Source: arxiv.org/abs/2606.07519
#LLMResearch #CodeAI #NLP
29
<p>The days of learning a skill once and coasting on it for life are over. The new reality, according to Campus founder and chancellor Tade Oyerinde, looks a lot like a New York City gym before summer.</p>
<p>Speaking on the first day of the 25th annual <a href="conferences.fortune.com/even…">Fortune Brainstorm Tech conference</a> in Aspen on Monday, Oyerinde posited that AI’s rapid pace of improvement has permanently changed the calculus of education—no matter where you are in your career. All that scrambling to deploy AI that companies and executives are doing right now? That’s not a one-time project that will suddenly reach a final resolution once and for all, said Oyerinde.</p>
<p>“I’m sorry if you’re exhausted,” Oyerinde told the audience. “You’re going to have to do that every year for the rest of your careers, ad infinitum.” </p>
<p>He predicts organizations will soon staff permanent “continuous learning, continuous development, continuous evaluation” departments, as standard as operations or finance. And with AI models now approaching recursive self-improvement—where each version helps build the next—the curve will only get steeper.</p>
<p>“The era of learn once and then you’re done for life is over,” Oyerinde said. “We’re going to have this exponential takeoff.”</p>
<p>Which brings us to the gym, another place where consistency matters. Just like “everyone in New York wants to be hot and fit for the summer,” and puts in two or three hours a week to get there, Oyerinde said, staying competitive at work will demand the same dedication and focus. </p>
<p>“If you want to be the equivalent of hot and fit in your career, you’re going to have to spend two or three hours a week learning how to use the most recent advances in AI.”</p>
<h2 class="wp-block-heading">A faster, smarter way to learn</h2>
<p>The reframing of continuous education as maintenance rather than a single milestone also applies to the traditional thinking around the way schools and curriculums are designed and built. Colleges and universities often take the same approach for vastly diverse student bodies. Instead, AI can map each student’s knowledge at an “atomic level” and route them through custom pathways, letting strong students skip ahead while others work on other gaps, explained Oyerinde.</p>
<p>The result, he claims, is teaching people “about five times faster.” That was the rationale behind Campus’s 2025 <a href="campus.edu/blog/press-releas…">acquisition of Sizzle AI</a>, a learning startup founded by Meta’s former AI chief Jerome Pesenti. </p>
<p>Hadi Partovi, who founded Code.org 13 years ago, recently renamed it to CodeAI to reflect the evolution in coding. The organization has focused on helping students learn the basics of computer science and the technology remaking their world. Now that mission has fully embraced coding and AI. </p>
<p>Every student, said Partovi, needs to grasp how AI actually works, then learn to build with it, and to use it responsibly. </p>
<hr class="wp-block-separator has-alpha-channel-opacity"/>
<p><em><strong>More from Fortune Brainstorm Tech 2026:</strong></em></p>
<p><em><a href="fortune.com/2026/06/08/anthr…">Anthropic’s Boris Cherny, creator of Claude Code, says there are days he manages tens of thousands of AI agents at once</a><br></em></p>
<hr class="wp-block-separator has-alpha-channel-opacity"/>
<p>“This is the most powerful technology that mankind has ever created, and anybody has access to that power,” he said.</p>
<p>Partovi described himself as a “cautious optimist,” but suggests not treating AI as “this magic thing that’s been created from above.” Instead, it’s something humans have developed, and everyone should help shape it. That applies to curriculum and learning to code, too. </p>
<p>However, just because AI can read, write, and do math, doesn’t mean schools will stop teaching these skills, noted Partovi. But the rote parts of coding, he said, like memorizing where semicolons and brackets go, no longer matter. What will continue to be important is computational thinking, logic, planning, and problem-solving.</p>
<p>“I also think we need to start questioning what students should learn,” said Partovi. “My guess is nobody here in this room uses calculus day to day, and no employer or almost no employer except maybe <a href="fortune.com/company/spacex/" target="_blank">SpaceX</a> or a few other places, but mostly nobody is hiring the calculus experts. But every student is struggling and thinking they need to get really good at it for almost no purpose whatsoever.”</p>
<p>And as education is pushed to evolve, given the way AI is reshaping learning and job opportunities, there’s a need for faster evolution in the way learning takes place.</p>
<p>“The gap between what’s relevant and what schools are teaching is growing as fast as the AI models are changing,” said Oyerinde.</p>
<p>And as students and companies adapt to the need for continuous new knowledge about AI, the thinking applies to those who are well beyond school or prime executive age. Karin Klein, founding partner at venture firm Bloomberg Beta, said AI can benefit all rather than a few. Klein participates in “talking circles” with icon Gloria Steinem, who is a lifelong learner.</p>
<p>“At 92, Gloria’s still learning,” said Klein. “I taught her how to use AI about a year and a half ago, so we all have no excuses, right?”</p>
<p>This story was originally featured on <a href="fortune.com/2026/06/08/ai-ad…" target="_blank">Fortune.com</a></p> fortune.com/2026/06/08/ai-ad…
127
Jun 9
code生成AIは、
破綻の決まった宗教と思って見てます。
もう見捨てたから、頑張ってな生成codeAI陣は。
いつかは分からんけど、
使用量の大幅な値上げがあって、
最終的な信者は使用料を支払えることに
快感を感じるようになると思う。
使い続けろよ。
借金してでも、破産してでも、
code生成AIを。
1
1
25
A bold and timely evolution that reflects where technology and education are headed, but how will CodeAI ensure students learn critical thinking alongside AI skills?
6
Code.org is now CodeAI. Students need digital fluency to thrive in a world reshaped by AI. Learn more: code.org/en-US?utm_source=tw…
1
11
9
1,595
Jun 4
Our long-standing partnership with @codeorg (now CodeAI) has expanded access to AP Computer Science Principles for students of all backgrounds. As AI reshapes the workforce, we too believe students need foundational skills and hands-on exposure to the future of work in high school to succeed in an AI-driven workforce. Learn more about CodeAI's renewed commitment to preparing students for the future.
1/ Today, Code.org becomes CodeAI.
AI is reshaping every part of students' lives before any school has decided what they should understand about it.
2
10
9
7,614
Jun 3
سایت Code.org بعد از ۱۳ سال فعالیت در حوزه آموزش علوم کامپیوتر، نام خود را به CodeAI تغییر داد.
تمرکز دنیای امروز از «کدنویسی محض» به سمت «هوش مصنوعی، علوم داده و سواد دیجیتال» حرکت کرده و هدف، تربیت نسلی است که به جای صرفاً استفاده از تکنولوژی، توانایی درک، جهتدهی و خلق با AI را داشته باشد.
Jun 2
Code.org is changing its name.
13 years ago, Code.org launched with a simple idea: every student should learn computer science — to learn how technology works and how to create it — not just how to use it.
After more than 2 billion hours of learning, the focus of CS has moved from coding to AI. As AI reshapes every part of daily life, students need digital fluency: the ability to understand AI, direct it, question it, and create with it — built on the foundations of computer science, AI science, and data science.
Today, Code.org enters its next chapter as CodeAI.
Our curriculum, teacher training, and frameworks, are there already: AI Discoveries and AI Foundations are free and in classrooms now. The K-12 digital sciences pathway is expanding.
Our goal is a generation with agency over the systems shaping their lives — prepared to shape the work, civic life, relationships, and meaning that come after.
Welcome to CodeAI.
2
102
May 29
Hiring ‼️
Web3 Sales Representative
Company: CodeAI a.s.
Pay: $60k - $120k / year (Up to 20% commission on completed deals)
Work type: Remote | Flexible Cooperation
Core Requirements:
- Proven experience as a closer, financial advisor, or sales professional (experience in crypto, forex, or investments is a major plus).
- Strong communication, negotiation, and high-impact sales skills.
- Ambitious mindset looking for long-term cooperation and the ability to build/recruit your own sales team.
What You'll Do:
- Drive sales and close deals for an international project in financial technology and online investments.
- Leverage a modern automated CRM system to track leads and manage your sales network pipeline.
- Collaborate with affiliate partners and direct sales teams to scale brand reach at an early project stage.
How to Apply:
Send your sales background, track record, and contact info to:
✉️ Codeai.jobs@gmail.com
May 29
Will be dropping more remote roles in the evening
:
Stay Tuned in
📌
3
3
8
2,722
Web3 Jobs (Remote)
Sales Representative
💵 Salary: $60k - $120k
at CodeAI a.s. Apply 👇
web3.career/sales-representa…
1
15
1,108
May 11
שמח לעדכן ששלושת ההרצאות מהמיטאפ CodeAI שלנו זמינות.
ההרצאה הראשונה היא של אנדריי מ-intuit שמסביר לנו איך מודלי שפה יכולים ליצור קוד בסקייל דרך Genetic Algorithm
youtube.com/watch?v=c86r2RsF…
1
23
2,795
Apr 26
תזכורת, המיטאפ שלנו, CodeAI, מתקיים מחר והוא יכלול 3 הרצאות מעולות על הדרך אפשר להשתמש ב-AI לכתיבת קוד.
החל ברמת המפתח/ת ועד לרמת החברה. ממליץ בחום.
5
29
17,349
Apr 23
אני עושה הפסקה מגוגל כדי להזכיר לכם שיש לנו תאריך חדש לאירוע CodeAI. מי שנרשם שלחנו לכם מייל לאשר הגעה (אנא עשו זאת) ומי שרוצה עוד להירשם יש עוד מקום.
לצד Intuit יהיו דוברים מ-2 חברות מאוד מבטיחות (אפילו היו ברשימה) שידברו איך אפשר לכתוב קוד בצורה טובה יותר עם AI.
3
19
1,412
Apr 22
Another Insights from Zhihu contributor Kaiyuan💻
📝My Checklist for K2.6: Architecture → Hype → Experience → Training Methods
🏗️ Model Architecture:
• Same as K2.5 (native multi-modal, large params) — proves the backbone still has potential
• AttenRes/KDA likely saved for the next major version.
📣 Hype & Real Performance:
• Leads in Coding, long-horizon tasks & AgentSwarm ability 🚀
• Key win: Engineering closed-loop capability — maintains architecture consistency, iterates with tool calls, mimics senior engineers (diagnose → implement → verify → optimize).
• Real-world wins: 4000 tool calls, 12h continuous iteration → throughput jumps 15→193 tokens/sec (Qwen3.5 deployment); boosts exchange-core throughput.
• Strong in Python/Go/Rust, frontend, DevOps — perfect for complex R&D.
• Goal: Lead open-source Agent models (24/7, collaborative, task-executing) — not just "better at coding."
🧪 Quick Test (For Fun!):
• Multi-modal space composition test failed (both fast & slow modes) — model is too popular, needs a membership! 😂 (Unreliable test, just for laughs)
🔮 Training Guess:
No official details, but RL weight likely increased in SFT RL hybrid paradigm (drives coding/Agent gains).
Final Thought: Chinese labs (Moonshot, Zhipu, ByteDance, Alibaba) are closing the gap with global leaders — impressive progress with limited resources! 💪
#KimiK26 #LLM #AI #OpenSource #AgentAI #CodeAI #Tech
🔗Full article (CN):
zhihu.com/question/202971452…
Apr 22
Kimi K2.6 Open Source Release Breakthrough 🚀
Evaluation & insights from Zhihu contributor toyama nao 🧐
✨Short Take: Every model shines in its own era! ✨
💡Core Lowdown:
• Moonshot’s "Lunar Exploration Program" levels up! 🚀 K2.5 = unmanned lunar orbit; K2.6 = manned lunar orbit — way more human-like in real-world tasks & long-horizon missions.
• Reasoning ability hits #1 among Chinese LLMs, snatching back the top spot from Seed! 🏆
• Price creeps up slightly: 21 → 27, but total cost stays in check (similar chain-of-thought efficiency).
• Non-reasoning mode struggles: Simple tasks = okay (≤10K tokens); complex tasks blow up to 20K–30K tokens (15K cap → nearly half ungraded).
🔍 Key Improvements vs K2.5:
✅ Long-chain Reasoning: Fixes K2.5’s incomplete solution exploration; fully checks all possibilities. Tradeoff: 3x more tokens for complex logic vs K2.5 (GPT-5.4/Opus 4.6/4.7 use 1/2–1/6 the tokens!)
✅ Coding Ability: Agent programming leaps — outperforms Sonnet 4.5 in regular frontend/backend dev (now "production-ready"). Weak on niche bug detection (fixed in official K2.6? TBD).
✅ Output Polish: Finally wipes out NBSP spacing bugs (only 1 case found!). Keeps outputs concise — tokens spent on logic, not fluff.
⚠️ Flaw:
Hallucinations unchanged from K2.5 (mid-tier). 2x more tokens spent on self-verification, but it’s useless for long-text retrieval/summarization.
🔗Full article (CN):
zhihu.com/question/202971452…
#KimiK26 #LLM #AI #OpenSource #AgentAI #CodeAI #Tech
3
27
3,463