sunil mallya

@dandigangi @QuinnyPig Somehow they didn’t learn that compliance has standards, apparently they are going through their auditors for standards!

@sunilmallya @QuinnyPig bro isn’t even 30 and thinks he could write a book on compliance already 😩

Absolutely none of this would have been done if they hadn’t gotten caught not doing it. “Untrustworthy” is an instant dealbreaker for freaking compliance. You do not lie to your auditor, full stop. Much less your customers.

@ypatil125 @abhijaymrana How are you different than an Anthropic + FDE?, it seems that your platform too needs FDE embedding to solve use cases specific to an enterprise

@abhijaymrana Small clarification: What we do at Applied Compute is much broader than “RL-as-a-service” We’re building the platform enterprises use to create high-quality agents specific to their business. RL is one tool we sometimes use to improve those systems.

There were a few RL-as-a-service companies that emerged this year (with one recently acquired by DoorDash for a very generous price). Some aspects of custom post-training are exciting, but the economics seem tricky here. Fundamentally, the RLaaS thesis is that enterprises should post-train models in-house to: 1. Outperform base models on uncommon, domain-specific tasks. - For DoorDash, this includes optimizing ad relevance and order recommendation models. Enterprises at DoorDash's scale have petabytes of proprietary data, and are evidently willing to spend hundreds of millions for talent alone. - Tangentially, these complex domain-specific tasks also make for great training data, which is why some of these companies make most of their revenue from data sales, not enterprise RL. - This is usually only necessary if the model is the bottleneck, which isn't the case for almost all agent deployments. Context/harness engineering go a long way. However, if you have you have a measurable objective + tight feedback loop (like Decagon), improvements start to become nonlinear and significant. 2. Derisk from frontier labs (long-term). - Owning your own model insulates you from OpenAI/Anthropic API pricing, which is why Cursor is racing for a SoTA coding model. - Regulated industries may also prefer local models for data security (especially if training truly becomes commoditized and they can justify the spend). In reality, as Michael mentions, there are ~zero businesses that are immediately ready for RL. Teams like Applied Compute spend most of their time transforming data and mapping processes (which can spiral into months of work) and even building agents for the enterprises before they can train. This feels slightly distracting—closer to McKinsey than OpenAI—but is also the only way to get AI in businesses today. But is this a good business? In some ways, it feels too early and hard to commoditize. Michael's write-up seems a bit frustrated. Also, there are a few existential questions here: - Is a six-figure training run + continual improvement costs worth it for most businesses? - It's almost like (beyond compute) hiring is the bottleneck to scale here. Sure, AC is innovating on RL infra but why won't businesses hire AI talent internally instead of paying AC millions on top of already high training costs? This doesn't seem sticky. - If Opus 5 is substantially better than Kimi K2.5 (which is already distilled from Opus 4.6 lol) does the fine-tuned model become obsolete? Also, the cost of the training run is amortized across the time period between runs (e.g. Cursor has to "pay off" Composer 1 in the period between launch and the Composer 2 release), which applies additional pressure. - Or, if they re-train on the next open-source model, will any of the initial post-training investment transfer? Maybe the data doesn't have to be re-processed, but unsure if training is cheap enough for this to be dismissible yet. Also, part of the reason these are still open questions is because the impact of post-training is difficult to quantify—there's no reliable way to attribute business outcomes to post-training improvements. - Sure, you can create a benchmark to track model improvement on some tasks. - But there's still a gap between [progress on an ad-recommendation benchmark] and [direct revenue growth for the business] that is latent (and hard to measure, given the lack of a single variable in a company's strategy). Realistic benchmarks are still an unsolved problem. In-house post-training is a no brainer for some businesses. Maybe DoorDash will benefit, and it's obvious Cursor and Decagon will. But broadly, I'm unsure how large or sticky this market really is.

@PrajwalTomar_ No it doesn’t make LLM smaller, just the kv cache

Google just dropped TurboQuant, and I'm about to get so much more out of my Mac Mini now lol. It makes LLMs 6x smaller and 8x faster with zero quality loss. Now I can run insane AI models locally for free. → Bigger context windows → Way faster processing → Completely secure Google could have kept this to themselves. They didn't. Huge respect for pushing the entire industry forward. We're just 3 months into 2026 and so much has been happening already. If you're not paying attention, you're already behind.

Dear @nikitabier and @Support, why do genuine posts don’t get the visibility they deserve? I can’t even reach the people who follow me! This has been a consistent observation for many posts over the last few months, despite being a paid user.

@GergelyOrosz I always wondered how people with barely any real work experience build compliance products that require such nuanced knowledge, lots of trust and painful discipline that you learn from experiencing the flows.

Damning evidence suggesting that compliance certificates issued by Delve (a startup founded in 2023) are fraudlent + worthless I never understood how eg Cluely could be GDPR, SOC2, HIPAA compliant in ~a week. Now we know: they probably aren't. Just wild substack.com/home/post/p-19…

@levie Two pizza model for the win

“AI exposed jobs may increase hiring and attract higher wages. It all depends on a) elasticity of consumer demand and b) number of AI exposed tasks in a job.” This is a key point. We’re going to see lots of AI automation emerge that has the opposite effect that we expect, because the cost of doing something goes down and greater demand for that service exists at lower prices. Take a *very* simplistic example in agentic coding to see what happens when you can dramatically increase output per $ of engineering budget. Before AI, a mid-sized company or team within a large company has a project they want to build software for. It takes 50 engineers to fully resource the effort, but the project doesn’t provide the ROI to fund it compared to other initiatives. Or the company knows its expertise isn’t in building software so it’s not even worth starting. So they hire 0 engineers, and don’t start the project. Now, AI agents make it possible for this to be a 10 engineer problem. All of a sudden the ROI calculus immediately changes on starting up the project. So now instead of hiring 0 engineers to do the project, the company hires 10 with AI agents. This has endless implications in coding, in particular, because coding can now have impact for anything from doing internal workflow automation, systems integration, data analysis, as well as customer-facing product innovation. By bringing down the cost of writing code, we can just begin to use it for far more. This will likely play out in a number of other job families as well, where lowered costs or higher output will lead to more demand. Now, not all of this will be smooth. For instance, there may need to be some reallocation of talent across the economy to move from some places of excess supply to places of lower supply. This could be bumpy at times, but the dynamic holds.

@anishmoonka We can’t see the decision that coding agents make for us, that’s why I built and open sourced Agentdiff to help developers including myself gain visibility. github.com/sunilmallya/ag…

Amazon had four Sev-1 outages (their highest severity level) in a single week. Internal memos say AI-assisted code changes were a contributing factor. The timeline here is wild. In October 2025, Amazon laid off 14,000 corporate employees. In January 2026, another 16,000. That’s about 30,000 people in five months, roughly 10% of the corporate workforce. CEO Andy Jassy said the cuts were about culture, not AI. During those same months, Amazon set a target: 80% of developers using AI coding tools at least once a week. They tracked adoption closely and blocked rival tools like OpenAI’s Codex. Even so, 30% of developers still hadn’t touched Amazon’s in-house tool Kiro by January. In December 2025, Kiro caused a 13-hour AWS outage. The AI tool had production-level permissions and decided the best fix for a bug was to delete and recreate an entire live environment. A second incident involved Amazon Q Developer, another AI tool. Amazon blamed both on “user error, not AI.” But quietly added mandatory peer review for all production access afterward. Then March 5: Amazon’s retail site went down for about six hours. Over 22,000 users reported checkout failures, missing prices, and app crashes. Amazon called it a “software code deployment” error. Five days later, SVP Dave Treadwell made the normally optional weekly engineering meeting mandatory. His memo acknowledged “GenAI tools supplementing or accelerating production change instructions, leading to unsafe practices.” These problems trace back to Q3 2025. Amazon’s own assessment: their GenAI safeguards “are not yet fully established.” The new rule: junior and mid-level engineers now need senior sign-off on any AI-assisted production changes. Treadwell also announced “controlled friction” for the most critical parts of the retail experience. For context, Google’s 2025 DORA report found 90% of developers use AI for coding but only 24% trust it “a lot.” An Uplevel study of 800 developers found Copilot users introduced 41% more bugs with no improvement in output. Amazon is finding out what those numbers look like at the scale of a $500 Billion revenue company, with 30,000 fewer people on staff to catch the mistakes.

@fenestbuc Just built it for Claude code now, but the architecture is such that it could extend easily for other CLI tools. Just need to build those pre and post hooks

@sunilmallya this solves a real problem. when 90% of your codebase is agent-generated, git blame becomes meaningless without the reasoning context. attaching prompts and task context as git notes is the right abstraction. curious if this extends beyond claude code to other coding agents too.

Agentdiff: Git blame for coding agents New open source project to track claude code changes to track its prompt, reasoning, and task context. Agent metadata is auto-attached as Git notes, so PR reviewers get the full AI context without leaving git log. Audit AI-generated code like a human teammate. github.com/sunilmallya/ag…

@levie Dear Aaron, you build a cloud file system many years ago. remember?

@vasuman Hiring got 10x harder. That’s what it is

After interviewing CS students, one thing is clear: too many are drinking the koolaid that software engineering is unimportant because coding models are getting better. So they try to "differentiate" by pivoting to product, GTM, or CoS. Do not stray from the tech. Most founders, myself included, are looking for deeply technical people regardless of role. In a world where AI makes you 100x, it pays to know how to build agents that are robust. Every founder's fear is that your agents are brittle because you have no production software engineering experience. It's also easier than ever to learn. You can genuinely become top 10% in a week and top 1% in a month. The bar is on the floor because no one wants to learn it. I saw a tweet from a huge CEO saying "it's actually better to NOT know how to code." Holy COPE. If you're a student "pivoting from just SWE," chances are you're neither here nor there. It's not ok to be mediocre at GTM/Product while simultaneously being mediocre at SWE. And it's much easier to become an exceptional self-taught engineer in college than an exceptional self-taught GTM/Product person, which almost never happens without real business experience. Learn software engineering, you will be rewarded.

@nesbubuu @Alibaba_Qwen @ggerganov @claudeai It’s a 12pro. Should work fine on yours too

@sunilmallya @Alibaba_Qwen @ggerganov @claudeai I’m on a 13 and figured it would be too old what’s your setup?

The new multimodal Qwen3.5 @Alibaba_Qwen running locally on an old iPhone 12 with pretty good speed (10 tok/s) Uses a modified llama.cpp backend to support metal memory layout for Apple A series chip and SSM ops. Building on top of great library by @ggerganov and @claudeai code with zero iOS experience and old kernel dev experience :)

If you take a catch running back in a World Cup game you automatically win. Thanks @akshar2026

@scottastevenson This is great news for companies to review contracts faster :), reduce that sales cycle for me

We’ve raised $40m in addition to the $50m we raised last October. We’re seeing record-breaking growth in 2026, with lawyers booking 410 demos of Spellbook last week. We now service 4,000+ in-house legal teams and law firms in 80 countries. theglobeandmail.com/business/artic…

@JustinLin610 @JustinLin610 Great set of models, did some work on getting them run on old iPhone 12

final shot. these are the small models that i told u!

@DeepLearningAI I got it working on an old iPhone12 :) x.com/sunilmallya/st…

Liquid AI released LFM2.5-1.2B-Thinking, a 1.17-billion-parameter reasoning model that runs in under 900 MB of RAM and operates about twice as fast as similar models. Designed for small devices, it performs competitively on reasoning benchmarks and is suited to agents that orchestrate tools, extract data, or run local workflows without cloud compute. Find all the details in The Batch ⬇️ hubs.la/Q045gVzd0

cc @Scobleizer more local model goodness, this time cracking the vision + text on Qwen3.5

*some tech notes* llama.cpp crashes on iOS (older iPhones), as it uses 16K system page size vs 4K Metal requirement Qwen3.5 uses a hybrid SSM/attention architecture (delta-net). Metal GPU kernels for SSM ops produce garbage output. Pure CPU works but is slow. Since iPhone12 has unified memory for CPU and GPU, I'm able to offload select operators on CPU, so no PCIe copies, just pointer handoff. Once llama.cpp has full support for the operators, it'll get even faster!