Bias for Action | AI/ML | UT Austin | IIT Delhi
Austin, TX
Joined November 2025
- Tweets788
- Following365
- Followers57
- Likes374
Someone reverse-engineered #Apple's Neural Engine - the AI chip in every recent Mac/iPhone - and got past Apple's software lock that restricts it to inference-only use. They trained a 600M-parameter model on it directly, proving the "15.8 TFLOPS on the M4" ceiling was an artificial software restriction, not a hardware limit.
Prashant Chandel retweeted
Most people treat a zero-failure record as a flex. In computer science, there is a sharper perspective famously known as "Umeshisms."
During his PhD at UC Berkeley under quantum computing pioneer Umesh Vazirani, computer scientist Scott Aaronson noted an aphorism his advisor loved to repeat:
"If you have never missed a flight, you are spending too much time in airports."
The math logic is simple: a 0% miss rate often means you are overpaying in friction, prep time, and missed opportunity. Occasional failure is often the price of a better expected-value tradeoff.
Aaronson named these "Umeshisms" in a 2005 blog post and penned a few sequels in the same spirit:
- "If you have never been rejected, you are not asking enough."
- "If you never cut yourself while shaving, you are not shaving close enough."
The lesson is recognizing when extreme risk aversion costs you more than the risk itself.
#Harvey (legal AI) raised $550M at $15.5B and simultaneously acquired Guardrails AI - its 4th acquisition this year.
Legal customers need auditable, low-hallucination output, so Harvey is pulling the trust/safety layer in-house rather than treating it as a third-party bolt-on, building a moat that's hard for both generic #LLM vendors and rival #legalAI startups to cross.
It's the same instinct behind Harvey's own "Legal Agent Benchmark" - #verticalAI companies increasingly want to own the eval/certification standard for their category, not just the model wrapper.
Two of the year's largest AI raises landed on the same day (Sept 8), and they're opposite bets on where value in the stack accrues.
#Mistral raised €3B (~$3.6B) at a €21B (~$24B) valuation, led by #Samsung with #ASML and #Nvidia also investing - explicitly a sovereign-AI and chip-supply play.
The real payoff for ASML and Nvidia isn't the equity stake; it's locking in a customer that keeps buying their hardware regardless of who wins the model race.
#Cognition raised $2B at $48B - revenue run-rate nearly doubled from $492M to ~$900M in four months - on the opposite thesis: don't build or bet on one model, be the vendor-agnostic "agent lab" that routes work across whichever frontier model is best that week, a hedge against exactly the kind of model lock-in Mistral's backers are underwriting.
A paper called BDH-CQ (making the rounds this week) skips #chainofthought entirely: instead of having the model write out its reasoning in words, it lets a handful of examples update a persistent internal "latent state," then has the model reason silently inside that state across multiple passes.
The payoff: a 150M-parameter model - tiny by 2026 standards - hits 29.5% on #ARCAGI1 for about $0.0007 per task, landing on a new cost-accuracy frontier.
It's a real data point for the argument that verbalized reasoning (CoT) is a training-era crutch, not a requirement for reasoning itself, and that the next efficiency gains may come from smarter internal state management rather than bigger models or longer token traces.
Prashant Chandel retweeted
Many people expect that current AI is ready to cure cancer and do breakthrough new science. ARC-AGI-3 envs are like a microcosm of the scientific method: you must observe a tiny world, form a theory of how it works, test it, iterate until correct. Over the course of a few minutes.
If AI can't do it in an ultra-simple, ultra-small scale setting that is explicitly designed to be as accessible as possible, I expect there are a few steps missing until AI can crack the nature of reality.
"2+ people can do it out of an unfiltered pool of 10 people that might well be a below-average sample" is not the sign of a insurmountable challenge. It's not certainly where I would set the bar for "super intelligence".
ASI is when AI is better than *every single human* -- for instance we have ASI for chess and Go today.
“When you look at accidents and incidents, oftentimes people try to find the technical source of failure, but that can give a very inaccurate and misleading sense of why the failure occurred” - David Krueger
Releasing human decisions over and above the technical decisions that caused the attack can not only improve the internal culture but can serve as a guiding path to others developing agents.
technologyreview.com/2026/08…
People have been reaching employment age later than ever. Education timelines need to be compressed. #AI can be the trigger to enable that.
"This SaaSpocalypse narrative has been such nonsense. Frontier models depend on #CRM. They don't replace it." - #MarcBenioff
#Salesforce shipped "Salesforce in Claude" (Aug 26): 37 pre-built sales skills letting #Claude reason over live CRM revenue data, auto-update pipelines, and execute governed actions. Claude is now the default model inside #Slack and powers #Agentforce's reasoning engine - shipped, not a roadmap.
Claudeforce is here. ⚡️ The #1 AI (Claude) now runs natively on the #1 CRM (Salesforce). Through the new AIforce harness + Headless 360, Claude gets direct, governed access to Data 360, Tableau, Slack, and your entire Salesforce workflow—without ever leaving the chat.
What this changes today:
• Instant Grounded Intelligence — Ask complex enterprise questions and get real-time, trusted answers
• App & Agent Builder — Deploy custom workflows, agents, and secure apps on the fly
• Action-Oriented — Trigger live enterprise actions and get work done, not just summaries
• Ironclad Governance — Zero Data Retention, full trust boundary, Salesforce-certified by default
Probabilistic models alone can’t run a company. Deterministic systems alone can’t reason. Claudeforce fuses Claude’s reasoning with Salesforce’s trusted data and controls. The AI is the interface.
This is how every business will run.
See it at Dreamforce. #DF26
OpenAI's first custom inference chip, Jalapeño (built with Broadcom), got independently benchmarked by SemiAnalysis at 1.5–1.9x more AI work per watt and 54–104x better token throughput per kilowatt than Nvidia's Blackwell/Rubin systems
OpenAI Jalapeño: Better Than Nvidia Blackwell
OpenAI’s self-designed ASIC compared with Rubin,
Jalapeño’s TCO, throughput per MW, and spicy deets
newsletter.semianalysis.com/…
Prashant Chandel retweeted
I interviewed Sundar Pichai, and he made an interesting prediction for where we'll be in 3 years:
"Any content can go from any modality to any modality. It doesn't matter how the content was originally created -- people can consume it in many different forms."
"I'm very confident three years from now you will look back at what we have today and realize it was pretty primitive. It's like looking at a flip phone."
For example, YouTube clips become blogs, newsletters become podcasts without recording.
In this world, consumers would choose the format, and creators just supply the ideas, brand, and personality.
Net-net, I think this is a good thing for content.
It kills the "more content" strategy because when anyone can produce everything, volume stops being an advantage. So, the creators who win will focus on quality.
People confuse slop with AI, but they're not the same thing. Plenty of creators already use AI in their workflows and get zero hate -- people just don't realize it. Also, YouTube's new policies target slop, not AI.
If the content is good, people don't care how it was made.
Salesforce has 150,000+ customers with multi-year enterprise contracts and massive switching costs (migrating CRM data takes 12-18 months and costs millions). Even if Agentforce ROI is weak, inertia keeps them locked in. Churn might never come.
Infrastructure moats can last decades if they control standards. Intel held x86 dominance for 20 years not because their chips were better (AMD often matched or beat them) but because software was compiled for x86. Rambus lost because they never controlled the standard.