Tim Rädsch

this is the surest fire sign that we are indeed in a bubble everyone. perhaps a lot of bubbles simultaneously. i will show myself out.

You really think we're going to scale data labelers to AGI? Today, we release the largest public long-horizon dataset of human digital work. 600h of long-horizon AGI research across 3 months. 🧵(1/n)

Step 1: Fire up the workstation for a Nano Banana Pro 🍌 weekend project Step 2: Try the default toy example Step 3: Default Toy example is too sexy for the Content Safety Filter🤔 @OfficialLoganK Great work with AI Studio, the safety filters might need some work.

@mervenoyann This is common reality for most academic datasets.

even academic datasets are mislabelled I continue on fine-tuning SAM3, this time on an undersea dataset to make it domain-specific horribly mislabelled (see below) I'm so tired 😄 can't a girl get a nice instance segmentation dataset to demo fine-tuning SAM3???

Really excited to see some of my team's work in the Gemini 3 model card! A lot of text is produced by RL training, we can learn a ton by looking! This stems from our new pragmatic interpretability approach. We found this understanding by running an LLM on the data. Do what works

there’s ai slop and there’s ai kino. this is ai kino.

WE ARE SO BACK

apple sauce is an underrated condiment for chicken

@zanbuilds CODE

@rohitrango Classic

running baselines takes 10 times the amount of time as running our own method where it is _obvious_ that the baselines will not perform well

@henrythe9ths OKRS

I’ve been told our company shouldn’t exist. COVID wiped out 80%+ of our revenue overnight. 4 months later, we were back, and in under 24 months, we hit $100M+ revenue/year and got a new name: Super[.]com We started as a travel chatbot. Then we pivoted. Fast. Launched 5 new products. Raised $85M after hearing 150 “no’s.” Rebuilt the team. Rebuilt the brand. That journey became a Harvard Business School case study. But more importantly, it made me the founder and operator I am today. Today, I’m giving away our entire Operating Playbook and offering select high-impact 1:1 advisory calls on Intro, alongside the founders of @zillow, @Reddit and veteran VCs and experts like @andrewchen If you’re a founder or operator navigating pivots, growth plateaus, AI transformation or fundraising headaches, I’ve been there. I can help. ✔️ Scaling to $150M+ annual revenue ✔️ Fundraising from top-tier VCs ✔️ Building in public, gaining visibility & press ✔️ Growth strategy, team structure & operational excellence ✔️ AI transformation --------------------- 🚨 Want the exact Operating Playbook we used to go from near-zero to $150M+ revenue /year for FREE? • Like and share this post • Comment "OKRS" I'll send you our entire Operating Playbook — including the real OKRs, internal templates, and execution strategies that fueled our turnaround and built . These are the same docs behind a Harvard case study and our $85M raise. No fluff — just what actually worked. This won’t be public for long. Tag a founder who needs this. 📞BONUS: If you're looking for personalized, tactical advice — from growth to fundraising to AI — you can also now book a 1:1 with me. Link’s in the first comment.

Read the preprint 👉 arxiv.org/abs/2502.15563 I’m around all week—DM if you want to chat VLMs or benchmarks over ☕

Fun weakness facts 📉 • Gemini 1.5 Pro: 23.6 % on “2nd‑brightest” (Claude 3.5 Sonnet: 61.1 %) & >10th on “object facing camera”. • GPT‑4o: Only 43.3 % on point‑depth & misses “presence‑of‑others”. • Llama 3.2 90B: strong at spatial; weak on blur (30.2 %) & noise (24.8 %).

Heading to ICLR 2025 🇸🇬 in Singapore next week and will present our newest research at two workshops. Check out our work on benchmarking VLMs with domain-specific data -> arxiv.org/abs/2502.15563 More detail below.🧵

@henrythe9ths Funding

I recently exited my $150MM+ annual revenue startup that's raised $200MM in venture funding and discovered something shocking. The way 99% of founders build companies is fundamentally broken. There are 4 funding models, but ONE new model works best in today’s AI era. The traditional models are failing founders: • Venture Capital: Founders often end up with less than 10% ownership and often walk away with nothing personally, even if the company is worth "billions" on paper • Bootstrapping: Founders have to make large personal financial sacrifices and 80% fail within 18 months • Boot-scaling: Founders drain runway and bet everything on a scaling event that fails 72% of the time. However, a small group of smart founders are using a new funding model to build AI-native companies. These founders are reaching $4-6M ARR in a matter of months, and they own 90%-100% of the company. Some are even building $3-5M ARR businesses with zero employees using this exact funding model. So, after talking to 100+ founders, I created an in-depth 10-page guide sharing: • Head-to-head comparison of all four funding models with EXACT metrics • Founder ownership percentages, dilution, and liquidity timelines for each model • How AI has changed what's possible ($3-5M ARR with zero employees) • Expected revenue growth, profitability timelines, and liquidity events • Your probability of success with each funding model (both as a founder and investor) • The psychological reality nobody talks about (what it actually FEELS like) • The shocking difference in founder stress levels and happiness Want the complete breakdown and analysis? • Like and share this post • Comment "Funding" • Follow me so I can DM you

Most things in startups are clay not metal. You can make 1000 pots and keep redoing your work, and you’ll be great by your 1000th. Almost always better than the person who spent all that same time just planning their first pot.

You have one life. Don’t settle for mediocrity.

👏look👏at👏the👏data👏you👏are👏training👏on👏

Just shake it off, shake it off! 🕺💃. .