Mo Patel

General take, exceptions ignored: This is a trap because low margin businesses cannot invest in technology. Their financial model doesn’t support it. Their investors have programmed in steady state returns. They cannot re-invest their low margin profits. This is where new entrants without the financial constraints, providing steady state returns to investors, can disrupt their businesses. You can see VCs shifting their attention to vertical AI companies that are disrupting low margin businesses for this reason. The other path is PE taking over these businesses and using debt to reinvest, reshape and repackage the business.

3) Harnesses make a huge difference in cost-performance. The very simple Pi harness (@badlogicgames) got the same success rate as harnesses from the LLM vendors with Opus and GPT 5.5, but at 2x less cost! Seems to be mainly due to smaller inputs to the LLM.

Bullish on harness engineering as the product interface, for enterprise use cases the question remains general purpose or domain specific harness wins out.

This is how AI forward companies are reshaping themselves

Beware AI FUD, many such cases.

We have seen in tech cycles where value accrues. Database companies have had the hardest time getting the combined valuations of the applications built on top of their product. As models become plug and play commodity with an optimizer sitting between the harness and model layer, value will move further up the stack.

@rasbt Wishing you peace in your recovery process. 🫶🙏❤️‍🩹

Usually, I post AI content and not personal life updates, but it somehow helps to get this off my chest. If I haven't been responding to messages, emails, or PRs, it's not because I don't care or don't want to. It's just physically hard at the moment, even with aids like speech-to-text. Life can be brutal sometimes. I was just getting back on my feet in March after a back injury that took 3 months to heal. I was so excited to be back coding and writing. And I had so many plans for this year. AI is a fast-moving field! And most importantly, I was hoping for this to be my last ER and hospital visit in life (or at least the next few decades), but no, life had other plans, and I have been dealing with a bad neck injury for the past 3 weeks. I have been seeing multiple doctors to get multiple opinions, and unfortunately, it may take 1–6 months until I can expect an improvement and start healing phase. So many talks and projects to cancel. And the worst feeling is the fear of missing out. Missing out on life and doing normal things like going for a hike on a beautiful spring day. And missing out on all the exciting AI developments and the missed project opportunities. I am working hard to get back on my feet as soon as possible or at least do some things even if it means I can do less and have to do things more slowly. I have done it before and can do it again!

@IanAndrewsDC Till Valhalla

Once more into the breach, dear friend, once more

@ChaitaliSuraj @trentjhughes Thank you! I love helping companies scale. 🌱↗️🌳

@trentjhughes @mopatel for Austin! He’s the best.

This one is a big one! So stoked to announce that Jason Pearl has been tapped to lead TRIBE Local. As we continue to grow, local chapters are the next evolution. Literally no one better to do it than Jason, an amazing leader who has built, advised and scaled a ton of great companies. Local chapters coming soon to the following: Dallas, TX Wilmington, NC Houston, TX Austin, TX Tampa, FL Los Angeles, CA NYC, NY Atlanta, GA Denver, CO Boston, MA Charlotte, NC And many more Interested? Let us know! Let's do this @jasonmpearl

@jaminball @Hammerspace_Inc Thank you! I read your blog post and was wondering if there was more. All good. Find your writing insightful.

@mopatel @Hammerspace_Inc This is the full interview. I don’t have a podcast, just filmed this video one off!

Checkpointing saves the state of an AI model during training so you can resume or recover without starting over. But writing massive checkpoints fast enough to keep GPUs busy is a real challenge. Below, David Flynn describes how @Hammerspace_Inc makes checkpointing efficient

@siicamore @patrickc Tripod Passing by Josef Chen, Routolotine by Kade Routolo are some examples. youtu.be/M0rY6gZ0ehQ?si… youtube.com/shorts/Hlg0vwt…

@mopatel @patrickc Interesting, do you have examples of these new techniques?

In which domains are elite practitioners celebrating the kids being better than ever before? Would love to read about a few instances. (Not just where there's one particular genius, such as Ashwin Sah's recent success, but where "the kids" as some kind of aggregate appear to be improving.)

Prepare to train your mind even more than the body. I have been in many matches where I was in great metabolic condition but my mind just didn’t work for variety of reasons and I couldn’t hear my coaches or I heard them but couldn’t act even the most obvious moves that I know very well. Here is how I have been successful and still working to improve: I have performed very well in tournaments where I was much more relaxed and casual. Instead of going on some type of a zone, I have tried to banter with my teammates etc. My coach asked me to visualize success and precise opening plays and not to veer from my planned path to submission. I have written down the path and checks on my phone which I am reviewing while waiting for match to start. Sure you can go 1 or 2 moves off the path but make sure to return to your game you have planned, visualized and trained 100s of time as opposed to the path your opponent wants to take which you may or may not be familiar with. I drilled my plan many times with variety of resistance and with partners that knew and partners that didn’t know. To work on the unpredictable, I do positional defensive and offensive sparring during regular training. Knowing that my defense is top notch gives me confidence to put myself in vulnerable positions offensively.

@julien_c Answers: 20 tks Research: 10 tks Coding: 100 tks Agent: 200 tks

Quick question, what's your expected minimum token per second when using a local LLM?

8 years since first @TensorFlow Dev Summit with @JeffDean kicking things off

@alvations Yes certainly challenging and the number of new approaches is staggering that it is hard to keep up, need a continuous updating survey paper

@mopatel We can theoretically have unlimited context length like RNN but due to hardware capacity there's only that much length we can squeeze into RNN limited encoder max-length

Is there such a thing as sub-seconds inference for decoder-only #llm for let's say input 1024 tokens? Expecting similar output length I'm getting ~3s throughput per payload on some "frontier" model and this is real tedious compared to calling old school #nlproc API

@alvations Is this helpful, I came across a talk from the team in YT recommendations and found their approach interesting github.com/LMCache/LMCache

This looks like some regressive move in terms of "blazing speed" inference we used to get for enc-decoder models. This used to be a no-go for any inference but UX stuns like "streaming tokens" has made users okay with the lag but impractical still for actual use.

@ronbodkin @RnaudBertrand @bgurley Agree with Arnaud. Directionally we should strive for safe unbiased open source models. However, the DeepSeek team needs to be commended for releasing a high capable model openly. Overall this moves the field forward and that is a good thing.

@RnaudBertrand @bgurley The alignment with CCP narrative is more deeply trained in. Yes you can fine tune it away but I’m not aware of proven ways to fine-tune a reasoning model while preserving its core capabilities:

All these posts about Deepseek "censorship" just completely miss the point: Deepseek is Open Source under MIT license which means anyone is allowed to download the model and fine-tune it however they want. Which means that if you wanted to use it to make a model whose purpose is to output anticommunist propaganda or defamatory statements on Xi Jinping, you can, there's zero restriction against that. You're seeing stuff like this 👇 if you use the Deepseek chat agent hosted in China where they obviously have to abide by Chinese regulations on content moderation (which includes avoiding lese-majesty). But anyone could just as well download Deepseek in Open Source and build their own chat agent on top of it without any of this stuff. And that's precisely why Deepseek is actually a more open model that offers more freedom than say OpenAI. They're also censored in their own way and there's absolutely zero way around it.

@jockowillink Till Valhalla! 🫶

No more watching the sunset with you on the deck. No more patrolling the perimeter. No more trips to the beach. No more sitting and staring and waiting for me to finish work. No more bounding in the snow. No more uncontrolled tail-wagging when it is time for a walk. No more sleeping at the foot of my bed. No more sniffing my face in the early morning to see if I’m awake. No more barking at the doorbell. No more chasing squirrels and rabbits (and bears!). No more scratching your belly. No more runs with with my buddy. No more big-eared silhouette waiting in the window when I come home. No more Odin. Thanks for being such a good boy and for standing the watch. I’ll miss you, buddy. YÜD BOY!