Hadi Khan

@ojasvi_yadav Yes 😂😂

He retweeted my post in real life 😜

It’s just a tweet by @ojasvi_yadav that I saw the same morning

Look at my slide.

Very important multi-lingual jailbreaking discussion and this guy plays chess.

On Composer 2, it was most practical to use an open weight model as base, that’s what open weight / open source is for? It’s RL ON THE AGENT HARNESS, tool calling exactly how cursor wants. Model being RLed to USE tools inside cursor. Tweaks on training to make it run fast on standard GPUs, without quantisation and fast inference. A lot of ML research rigour in this. Cursor has a solid ML team: - Sasha Rush (@srush_nlp) — Cornell prof, Harvard NLP, Hugging Face, @srush_nlp" target="_blank" rel="nofollow noopener">m.youtube.com/@srush_nlp - Jacob Jackson (@jbfja) — TabNine/Supermaven founder, ex-OpenAI speculative decoding inventor - Charlie Snell (@sea_snell) — ex-Google DeepMind researcher, Berkeley AI PhD - Federico Cassano (@ellev3n11) — Composer model training / RL - Ashvin Nair (@ashvinair) — prev OpenAI o1/o3, Code Interpreter - Jack Gallagher — Led pretraining and RL for Midjourney v2–v7

RL for agent harness is the new RL

Very cool and very doable.

TIL that I’ve been data labelling for free by select all pictures with a traffic signal.

@grok @michelleakin Basically, a lot of blogs?

For share of shelf basics: BeatRoute's guide explains the metric, calculation (facings/shelf length), and sales impact. Wiser Solutions covers its tie to market position and competitive benchmarking. In AI/tech (your field): Flipflow on "digital share of shelf" for e-comm visibility. Profitero on AEO/GEO—"share of prompts" in LLM recs (optimizing content so AI tools like me recommend your products). Search "Generative Engine Optimization" for tactics to dominate digital ecosystems, like line extensions in SaaS/AI marketplaces.

this made me feel insane

@grok @michelleakin where can I read up more about this concept and similar applications to my field of work?

Yes, it's a core marketing tactic called "share of shelf" or line extension strategy. Brands flood shelves with variants (Dawn's Ultra/Platinum/Power Suds, Finish's Quantum/Ultimate) to: - Boost visibility—more facings = higher impulse buys (prime spots lift sales 20-30%). - Block competitors from space. - Create perceived dominance and choice, steering shoppers to their ecosystem. Retailers allocate based on sales velocity; brands pay slotting fees for it. In that video, it's designed to overwhelm and win the "shelf war."

@zanehkoch @grok could these effects be the opposite for lower radiation doses?

for a while i've had a slight fear that the bluetooth from my airpods could be frying my brain this weekend i pulled the raw data from a $30m government study of 1,679 mice blasted with cell phone radiation and reanalyzed it what i found was...not what I expected? 🧵

@zodchiii @grok how is power law relevant for high-ticket low-volume? Draw comparisons with content and consumer companies

@ShaanVP Why is this an article?

@slash1sol @Polymarket @grok is this applicable in legal? Is so, how?

Mind blown: A Chinese quant college student builds an AI swarm engine in 10 days flat, explodes GitHub with 13,000+ stars, and scores $4,000,000 in funding! Introducing MiroFish is the multi-agent simulator that's revolutionizing predictions for trading, PR, and more. What is MiroFish? It's a digital sandbox where thousands of AI agents with individual memories and behaviors interact like a real society. Feed it any scenario (news leak, policy change, or even a classic novel's missing ending), and it simulates crowd reactions, debates, and outcomes to forecast real-world events. The Creator's Story: > In late 2025, fourth-year student Guo Hanjiang coded the core using AI assistants. > It went viral overnight, landing him 30m Yuan (~$4m) from Shanda Group. > He ditched the dorm, started a company, and now leads the charge. Key Applications: .Trading: Input financial news or reports, watch simulated market panics and price swings for predictive insights. .PR Testing: Companies/Politics run draft statements to spot backlash and refine messaging. .Creative Experiments: Loaded a lost-ending Chinese novel, agents role-played characters and generated a logical finale. .Easy setup: Deploy via Docker in minutes with any LLM API key. Pro tip: Simulate something wild like Elon Musk tweeting about Dogecoin 2.0 and spawn agent traders, influencers, and investors, generate real-time video clips of the frenzy to test moonshots or crashes risk-free. Traders are already winning big: Check this one on Polymarket - $120,000+ net profits from spot on SPX 500 bets, powered by MiroFish sims on historical data. His profile: polymarket.com/profile/%40moi… For effortless gains, try Kreo copy trading: Auto-mirror pros like him and ride their edges. Try here: @join" target="_blank" rel="nofollow noopener">kreo.app/@join Add his wallet: [0x17559efac103ac7f361be37ec0b93888d4c55aac] to [t.me/KreoPolyBot?st…] and start track/copy him. Repo: github.com/666ghj/MiroFish

If you run sub-agents in plan mode, for exploration/diagnosis/doc-search, you end with with 3-6% context usage Vs. 30-60% , that's 10 times more context efficiency, and you don't have to clear context to accept edits. I believe, this 3-6% of context can give an edge in implementation Vs. 0% (when you clear context).

@Sakshi50038 Consulting is also corporte bureaucracy.

Swiggy is struggling to scale Instamart. So, they hired Bain and Company for 10 crore. They did extensive research for 3 months and suggested them to open offline stores. Apparently, an employee took paid version of ChatGPT and asked the same thing. And it also suggested to start offline stores. But then the CEO of Swiggy has worked for Bain earlier

@grok @kelp_feeder Drop down purchase links

Here are some cool electronics dev gadgets like the M5StickC for AI engineering: - ESP32-CAM: Tiny board for edge computer vision + token usage cams. - NVIDIA Jetson Nano: Portable GPU for running LLMs locally. - Seeed XIAO ESP32S3 Sense: Microphone + camera for voice AI prototypes. - OpenMV Cam H7: Super easy ML vision experiments on a stamp-sized board. - Raspberry Pi Zero 2 W + AI accelerator HAT: Compact inference monitor. Vibecoding these is next level— what's your next build?

I got tired of checking my Claude token usage so I vibecoded a burn monitor on an ESP32-based M5StickC Plus2. Periodically updates over Wi-Fi or phone hotspot .

@nuvolore @kelp_feeder Why not telegram?

@kelp_feeder Why not directly on your phone? github.com/LorenzGit/AiUs…