Happy Bear

[Graham Neubig @gneubig]

This Fall at CMU we're teaching a new course on AI Agents!

The goal is that you learn how to create a scaffold, build evals, and train an agentic LLM using RL.

We'll try to balance theory and practice, and introduce modern frameworks and best practices.

[Mario Nawfal @MarioNawfal]

🇧🇷 A Brazilian amateur runner hits the awards stand and starts peeling off T-shirt after T-shirt.

Valdemar sold his chest to 11 brands, and he kept his word 😂

Writer: Sol

[极客杰尼 @seekjourney]

搞 harness 的朋友强烈推荐看看这篇 200w 浏览量的文章，Uber 内部扎实的工程经验。

如何多快好省的完成任务是一个挑战。

[​Տᥴαɾᥣᥱ𝗍𝗍 𝖵𝗂𝗌𝗂σᥒ🌐 @Scarlett1437561]

Baby ANIMALS WITH WEIRDEST ADAPTATIONS.

[ali @waterloo_intern]

I spent 48 hours with the Kimi K3 modeling code.

It took:
- 650 mg of caffeine (mandatory)
- 40 cans of LaCroix (optional... world record (?))
- 8 papers
- 6 months off my lifespan

Finally grokked the entire lineage of Kimi K3 and how we got here... every single step, since 2019 GPT-2

[Philip Kiely @philipkiely]

https://t.co/ZomEdoDLlJ

[Sam Hogan 🇺🇸 @samhogan]

We're releasing Inference AutoTune

Distill any frontier model into a 1-30B parameter task-specific SLM with only 25 lines of code

automatically route requests to reduce cost and latency by >90%

~2 hours and <$250 to train. You own the weights

Available in private beta today

[How To Prompt @HowToPrompt__]

China has open-sourced a sandbox that:

→ boots in 60ms (docker takes 200ms)
→ uses 5MB of RAM per instance
→ runs thousands of AI agents on one machine
→ isolates every agent at the hardware level
→ drops into your E2B code by swapping one URL

It’s called CubeSandbox, a hardware-isolated sandbox for AI agents that cold-starts in under 60ms with just 5MB of memory overhead.

It's a drop-in E2B replacement. Swap one URL. Keep your code. Kill the bill.

no Docker. no shared kernel. no cold start tax.

100% open source.

[Mrinal @Hi_Mrinal]

this was the best read on KV cache which was intuitively soo good to read about

https://t.co/fRWHUELkJ2

[Vivek Galatage @vivekgalatage]

If nothing else, this chapter alone should excite you!

https://t.co/yYrpFaCLZM

[h100envy @h100envy]

Alibaba engineer who leads Qwen explained the future of open agent models in 25 minutes - better than $2000 LLM training courses.

pre-train the base ->SFT -> RLHF -> tool use -> multi-modal -> ship a whole family (chat / VL / coder / math / QwQ).

That loop is why Qwen quietly became the most downloaded open model family on Hugging Face.

Qwen base + Qwen-VL + Qwen-Coder + QwQ reasoning - that's the stack.

Watch and save it, then read the article below.

[Thenationthailand @Thenationth]

Why Thais pay 300,000฿ for a 10,000 ฿ job
.
Why are people willing to pay a massive 300,000 baht bribe just to secure a Thai civil service job that starts at a mere 10,000 baht a month?
.
Following the recent shocking headlines exposing a massive exam corruption network, we dive into the cultural obsession behind landing a government position. While it seems like a terrible financial move on paper, the real value isn't the monthly salary. It’s a one-time investment for ultimate, multi-generational safety.
.
#ทุจริตสอบราชการ #สอบท้องถิ่น #ข้าราชการไทย #ThaiCivilServant #ThailandNews #TheNationThailand

[Oliver Prompts @oliviscusAI]

Turn Claude Code into a full academic research team with this guide.

> Deep Research: 13 agents, 8 modes
> Academic Paper: 12 agents, 11 modes
> Academic Paper Reviewer: 7 agents
> Academic Pipeline: 10-stage orchestrator with citation verification

Free for non-commercial use.

[Codez @0xCodez]

Claude Code creator:

“100% of my code has been written by Claude since Opus 4.5.

most of my coding routine is automated by loops & dynamic workflows"

In this 40-minute video, Boris shares how to build a self-improving agentic setup from scratch.

Loops + dynamic workflows + auto mode + routines - that’s the secret.

Worth more than a $500 vibe-coding course.

Watch today, then read below to learn how to build the same workflow.

[Alex Stone @a1exstone]

Monica Lam (Stanford Professor):

"AI writes shallow reports for one reason, you ask it ONE question. Our method asks dozens, like a journalist, and the same chatbot starts writing articles 25% better organized than top AI."

paper presented at a leading AI research conference, her Stanford lab (OVAL) unveiled STORM - a method already used by 70,000+ people to generate Wikipedia-grade, fully-cited articles on topics it has never seen.

the secret formula: 6–8 expert perspectives + cited expert interviews + ruthless outline + grounded section-by-section writing + blind-spot red team = a report you can actually trust.

watch the full breakdown to copy all 5 prompts into Claude.

save this post so the formula is ready when you need it.

[Brad @BradleyKellard]

Malcolm Gladwell revealed why you shouldn't go to Harvard:

1. America does not have a shortage of students who want science and math degrees. It has a shortage of students who finish them. Half of all high school seniors who intend to study STEM drop out by the end of their second year. The problem is not interest. It is persistence.

2. The obvious assumption is that smarter students persist longer. So Gladwell tested it. At Hartwick College, a small liberal arts school in New York, the top third of math SAT scorers took the majority of STEM degrees. The bottom third dropped out in large numbers. The data seemed to confirm it. Smarter kids stick around longer.

3. Then he looked at Harvard. The bottom third of Harvard's math SAT scores are equal to the top third at Hartwick. By the logic above, everyone at Harvard should graduate with a STEM degree. They are all brilliant. Nobody should be dropping out.

4. Harvard showed the exact same pattern as Hartwick. Top students graduated. Bottom students dropped out like flies. Even though the bottom Harvard students were objectively brilliant by any global standard. Something else entirely was driving the dropout rate.

5. That something is called relative deprivation theory. Human beings do not measure themselves against the world. They measure themselves against the people immediately around them. A Harvard student in the bottom third does not think I am in the top one percent of all students globally. They think that kid next to me keeps getting everything right and I keep getting it wrong. So they quit.

6. The research from UCLA puts a specific number on it. Your odds of graduating with a STEM degree fall by two percentage points for every ten point increase in the average SAT score of your peers. Choose Harvard over the University of Maryland and your chance of finishing a STEM degree drops by thirty percent. Thirty percent. Just to put a brand name on your resume.

7. Relative position matters more than absolute position when it comes to confidence, motivation, and self belief. The eightieth percentile student at Harvard looks up at the people above them and feels like they cannot compete. The number one student at a state school feels like they can conquer the world. That feeling drives everything.

8. The practical hiring implication is radical. Class rank matters more than institution name. Gladwell argues companies should have a don't ask don't tell policy for where someone went to college. Hiring only from top schools means missing the top students from every other school. That is not smart hiring. That is brand worship.

9. When choosing a college, never go to the best school you get into. Go to the school where you are guaranteed to be near the top of your class. Being a big fish in a smaller pond does not just feel better. It statistically produces better outcomes than being a small fish in the most prestigious pond available.

10. So why do we keep choosing Harvard over Maryland? Because we are flattered. Because the acceptance letter feels like validation. Because we make an irrational decision in a moment of enormous flattery and call it ambition. Gladwell's conclusion is simple and brutal. When we have the chance to join an elite institution we do things that are genuinely against our own interest and we feel great about it the whole time.

[Chrome @0xchromium]

Andrej Karpathy spent 2h showing how he actually uses AI day to day

he's a co-founder of OpenAI and led AI at Tesla, so when he shows how he works, it’s worth watching

and the whole session is just him telling the machine what he wants in simple terms, like he's briefing a coworker

watch what's actually happening the entire time:

> he describes the task in normal words
> it goes off and does the work
> he glances at the result and nudges it with one more sentence

that's the whole skill, and you've had it since you learned to talk

the only gap between that and a worker that runs on its own is handing that sentence a schedule and the tools to act

check his work, then build the version that keeps working when you stop

[Victor M @victormustar]

Before the week ends, let's acknowledge one of the most INSANE week ever for open AI, with 25+ notable open-weight drops across every modality:

🧠 LLMs

→ NVIDIA Nemotron 3 Ultra: 550B hybrid Mamba-MoE, only 55B active, 1M context, MMLU 89.1. NVFP4 variant claims ~5x throughput on Blackwell. First openly-weighted 550B hybrid Mamba-Transformer, closing the gap with frontier closed models.

→ Google Gemma 4 12B: fully open dense any-to-any (text/image/audio/video), 256k context, encoder-free, 140+ languages, AIME 2026 at 77.5. Shipped with a 23-checkpoint QAT wave (mobile ONNX + MLX). Most deployable model of the week.

→ StepFun Step-3.7-Flash: 198B sparse MoE VLM, ~11B active, SWE-Bench PRO 56.3. Apache 2.0.

→ Liquid AI LFM2.5-8B-A1B: edge MoE, just 1.5B active, 128k ctx, MATH500 88.8, MLX-ready. Best on-device option this week.

→ JetBrains Mellum2-12B-A2.5B-Thinking: their first open MoE, near-Qwen3-14B coding at 2.5B active. Apache 2.0.

🎨 Image gen (the surprise of the week)

→ Ideogram 4: their FIRST-EVER open weights. 9.3B flow-matching DiT trained from scratch. #2 overall behind GPT Image 2, top open-weight model on Design Arena + LMArena. Strongest open checkpoint for text-rich images, full stop. It has taste. Still can't believe this is open weights.

🔊 Audio & Speech (a breakout week for open TTS, 4 labs shipped)

→ Boson Higgs Audio v3 4B: 102 languages, 21 emotions, singing/whispering/shouting, sub-second TTFA.
→ RedNote dots.tts: the only fully continuous (no codec) open TTS pipeline, Apache 2.0.
→ Google Magenta RealTime 2: real-time music gen, <200ms latency, text+audio+MIDI. multimodalart ported it to PyTorch within hours with live ZeroGPU demos.
→ NVIDIA Nemotron-3.5 ASR: 600M streaming, 17x more concurrent streams vs Parakeet RNNT 1.1B.

👁️ Vision & VLMs

→ PaddleOCR-VL-1.6: SOTA document parsing at 1B params, Apache 2.0.
→ Baidu NAVA: 6.3B joint audio-video gen, best-in-class A/V sync, Apache 2.0.

🎬 Video, 3D & World Models

→ NVIDIA Cosmos3-Super: 64B omnimodal world model coupling action trajectories with video+audio gen, for Physical AI.
→ JD JoyAI-Echo: up to 5-min multi-shot text-to-video on LTX-2.3.
→ ByteDance Bernini-R + VAST TripoSplat (single-image-to-3D Gaussian splats, MIT).

[Dami-Defi @DamiDefi]

Realising Apple went public at under $2 billion and 15 times revenue in 1980.

SpaceX wants you to buy at $2 trillion and 100 times revenue in 2026.

That is not getting in early. That is being the exit for venture capitalists who have held this equity for years at a fraction of what you are being asked to pay.

Almost none of the retail investors buying this IPO will read the 300 pages before the book closes on June 11.

That is your entire competitive advantage right there.