Ivan Chong (張建生)

Looking at newly release open weight LLMs is like picking a new driver for my golf bag. Last week was pretty exciting as 3 frontier-level open weight models were released that actually fit on local AI setups. GLM 5.3 Flash. Qwen3.8-Flash-Next. DeepSeek v4 Flash Vision Exp. The community pounced. Tons of great serve script recipes, updates on serve parameters, updates with speculative decoding, updates with new quants... Benchmarks on prefill, decode, concurrency, context, SWE scores, tool calling, etc. Amazing work to get and post Day0 numbers. I'm currently running a proven serve script for DeepSeek V4 Flash 0731. It's reliable, robust and has done great work for me. It's my driver that gets me on the fairway in a good position to score. Now, I'm in a situation where I am tempted to go for the newest thing that promises better potential numbers - but would take a few rounds to adapt my swing. There's also the possibility that the wonderful benchmark metrics don't translate to my specific game. UNPOPULAR Opinion: It's a great problem to have. If I wasn't already really happy with DSV4-Flash-0731, I'd jump on these newly released models in a heartbeat. But I realized I have the luxury of waiting and watching to see if there's anything that offers true stepwise improvement for my game.

Truth

This is why it’s critical for the US (and NVidia) to develop best in class open weight models. I do not believe they will be available in the future. Both the US and the Chinese governments have reasons to restrict access.

Really fun watch- highly recommended if you’re a girl dad.

This car was designed with Tiger Woods in mind, amiright?

Good friend who works as an editor at ESPN developed this website using AI. This is the way.

Really nice summary on the brilliance of NVidia’s strategy. Achieving a dominant market position with the best AI chip on the market, the GB300, Jensen was clear about his competitive moat, the software ecosystem built on his CUDA platform. His appreciation of where his competitive advantage sat factors into his strategy going forward. It’s all about the ecosystem. His competitive threats are not only from other chip manufacturers like AMD, but also from his biggest customers (eg Google). Jensen countered by investing in his ecosystem, funding neocloud hyperscalers to cultivate and farm demand. Risk from demand concentration also came from the meteoric ascendence of frontier AI companies OpenAI and Anthropic. Jensen countered again with an ecosystem play - cultivating open weight LLMs. Open weight models not only take share away from Anthropic and OpenAI (thereby keeping them gently in check) but also diverts that demand over to the neocloud companies, thereby strengthening the earlier pillar on his ecosystem strategy. Open weight models also create opportunities for post training and fine tuning, workloads that create even more demand for NVidia GPUs… there’s layers to the synergy. Such an elegant strategy.

This is genius - offering customers a gaming experience in exchange for their efforts to collect data that you can turn around and monetize via AI.

Winning strategy for open weight models vs Anthropic and OpenAI - train base models for specific industries or knowledge domains and offer equal performance for far less cost. Because the weights are open and freely distributed, an entire ecosystem can be cultivated and applied towards purpose trained models.

I've been using local AI (GLM 5.3 Flash, in particular) to review contracts, legal entity formation docs, and partner NDA's. Do I still pay a lawyer? Yes, but my interactions are much different. I ask for legal strategy and review instead of legal legwork. My legal fees are going to be far more reasonable going forward.

Frontier models are great for unbounded or unscoped tasks but once a recurring tasks has clear specification, SFT and RL training for even the small models will outperform frontier.

Bull case for Ethereum

Heard something on the All In podcast that blew my mind… Anthropic advocates that AI provides “consent.” Excuse me, but there’s something super creepy about being required to ask for consent in prompts.

MLB using AI to identify the type of pitch a pitcher is setting up to throw based on pattern recognition on how they position their glove or set up their feet. AI will be used increasingly for competitive advantage in professional sports.

Ngl - X humbles me every day by putting banal posts on my timeline that get 10k views. I try and put thought into my posts about how AI or blockchain will change life as we know it. But a post about what is appropriate to tip on a $75 meal lands on my timeline with thousands of likes and tens of thousands of views and totally mogs anything I’ve ever posted. Some days it’s hard not to take it personally…

Reflection AI. US open weight model that makes it easy to train with your own data. It’s expected to compete against Chinese open weight models out of the gate but trail frontier models. Their strategy to compete against OpenAI and Anthropic is to encourage companies to use their own data to train Reflection’s model. I’m confident trained models can beat out frontier models on pre specified work. This is the way.

This was my experience. I used a local DeepSeek V4 Flash to fill out tax returns. I compared them with the filing my CPA charged me for filing. The differences were not material. Don’t try this with Cloud models.

AI + Ethereum blockchain This is the way

Wise words

If you ever wondered why Claude gives off such weird vibes…

Lately I’ve been using my AI agents to screen spam emails. Spam filters catch a lot of the junk but there are still many that get through. I forward cold intro emails to my agent and ask for an assessment. It’s amusing to see the agent highlight red flags and explain common spammer workflows. They also do followup replies to probe further - spam on spam, in effect. Makes me want to add voice and phone to my agents and unleash them on telemarketing calls.