bidhan

Being at the frontier - by the definition of it - means creating the frontier. You don't get to be at the frontier by following someone else. And creating the frontier often means discoveries that go against the established knowledge. We recently made such a discovery about distributed diffusion model training. A common way to optimize diffusion model training is by ensuring the numerical stability of their generation paths. We found that that's not true for the most efficient distributed diffusion model training architecture. We shared what works instead in our blogpost below. blog.bagel.com/p/stability-qu…

@kazdenc @bageldotcom 🫡

@bidhan @bageldotcom cadtech.fyi/directory/Bage…

in honour of Toronto Tech Week, my builder sunday project was tinkering with this Canadian Tech directory of companies and VCs 🇨🇦cadtech.fyi

this is indeed the default state we’re heading towards, which is, the owner of compute (openai, anthro, deepmind et al) decides where scientific progress will be. unless, the regular researchers can figure out how to coordinate distributed dark compute.

@rauchg 🫡

gm 🏃‍♂️

4 thoughts on early-stage hiring: 1/ If an engineer is trying to pick between a pre Series-B company and a BigCo/BigLab --> stop talking to them immediately. They are clearly not ready for a startup. 2/ If someone isn't willing to take a 70% cash paycut (relative to BigCo/BigLab) --> stop talking to them immediately. They will be unhappy/stressed. 3/ You learn a lot about a candidate during the negotiation/closing process. Do not be afraid to walk away if you get new information. 4/ Startups have zero work-life balance. If you are not willing to put in the hours, you are not in the right headspace to grind.

you cant stop claude

ironically, physical ai is the only part of ai industry where human researchers are absolutely required

@forwarddeploy @edenchan @amytam01 amazing event! thanks for organizing.

Chai + Samosas = wholesome time to catch up with folks and make new friends Thank you all for coming today!

as 80% of egocentric training data is generated in india, it’s safe to say that most robots will identify as indians

@SemiAnalysis_ imagine having a job title like “member of non-technical staff”

the biggest slur in Bay Area that words with starts with the letter N is "Non-T***chnical"

@rohanpaul_ai It's interesting because they don't even use NVIDIA exclusively. Amazon trainium and google tpus are a big part of their compute resources.

Anthropic drops a paper on the US-China AI race They believe the US and its allies may be able to lock in a 12-24 month frontier AI lead by 2028 if they close China’s access to advanced compute and copied model outputs. The report says China is not far behind because Chinese labs are allegedly using loopholes, smuggled chips, offshore data centers, and distillation attacks to stay close to US frontier labs. Anthropic frames compute as the central bottleneck of AI power, saying advanced chips are not just one input but the gatekeeper for training, deployment, revenue, experimentation, and future model improvement. The report says Huawei may produce only 4% of NVIDIA’s aggregate compute in 2026 and 2% in 2027, which is one of the paper’s sharpest claims about China’s semiconductor gap. Anthropic argues that distillation is systematic industrial espionage, because Chinese labs can use American model outputs to copy capabilities without paying the full training cost. The report claims a Chinese AI lead could enable automated repression, stronger cyber operations, faster military AI deployment, and broader authoritarian influence through cheap global AI infrastructure. Future frontier models may become a “country of geniuses in a data center,” meaning a single model cluster could act like a huge expert workforce for cyber, science, engineering, and military research.

having the dog in you is the biggest competitive advantage

@gauravisnotme generational arbitrage

You still have a week to shill your two-layer convoluted Anthropic SPVs on LinkedIn. Act Now.

@thsottiaux a twitterlike feed for all of my running codex agents to interact with each other

Now that the Codex app is close to being the super app. What should the super duper app do?

@k7agar this

true pre training for robot learning has not been done yet

@Dorialexander do you think “truly” open models should be crowdfunded?

money comes with strings attached

@shaanspatels @slashapp @jacob_bulbulia down 🫡

if you’re a founder in SF who loves poker and want to meet other cracked founders in the city come to the @slashapp office poker night hosted by @jacob_bulbulia and me on May 16th professional dealers, food + drinks on us Let me know if interested

@AnjneyMidha been thinking on this a lot. thought -> some latent representation of thought for agents feels like a more lossless approach.

thought to text is a very rich space to explore many teams across the ecosystem exploring this capability right now the best ones truly understand hardware software co-design can’t wait to see some of these approaches ship

.@DarioAmodei may I suggested using commodity hardware for your training?

@ClaudeDevs as always, competition is good for consumers. that's why we dont agi to be within just one company.

Usage limits are up, effective today we're: 1) Doubling Claude Code's 5-hour limits for Pro, Max, Team and seat-based Enterprise plans 2) Removing peak hours limit reduction on Claude Code for Pro and Max plans 3) Substantially raising our API rate limits for Opus models