Akshay Krishnaswamy

Chief Architect @palantirtech

[Palantir @PalantirTech]

Palantir honors America’s veterans and their families.

We are grateful for their service, their sacrifices, and the freedoms they helped defend.

To all who served, and to the families who support them: We thank you.

[Akshay Krishnaswamy @hyperindexed]

Link to the full white paper: https://t.co/yfXraz1YqN

[Akshay Krishnaswamy @hyperindexed]

Mitigating civilian harm is an essential part of AI-enabled military operations.

Great new paper by @PalantirPrivacy on how we're continuing to build foundational capabilities into the Palantir platforms for encoding critical doctrinal processes and ensuring auditability.

[sachin. @sachin2]

“Claude can help US kill more people more quickly ?”

Host: “With the help of LLMs, the US military has gone from hitting 1,000 targets a day to 5,000. That means Claude can help kill more people more quickly. Are you comfortable with that?”

Dario Amodei: “I am supportive of the US being more effective militarily. Having that ability be stronger doesn’t cause wars. It deters wars.”

“Do you want this country to be a more powerful actor on the world stage? I do. I’m a patriot.”

[Jawwwn @Jawwwn]

Palantir CEO Alex Karp on Americans' right to privacy: Most people don't realize how influenced they are by tech

"What they're often imputing is, 'I don't care what commercial companies know about me, because that's my choice,' which it is."

"But they're thinking of their opinions as being wholly their own, and at best, your opinions are a dialogue between your ideas and the outside consensus view."

"They're underestimating how much their thoughts are being influenced by the tech products they're interacting with."

"And because they underestimate that, they don't understand the in- the, the, the, the way in which they should be more concerned about privacy."

Via @WalterIsaacson @kscope_nyc

[Palantir @PalantirTech]

At DevCon 7, Palantir Software Engineers show that the biggest lever on agent performance isn't your model, it's your tools.

Watch what happens to speed and quality as you go from basic read/write tools to advanced capabilities like code execution.

The result: a mid-weight model outperforms a frontier model. Faster, at a fraction of the cost.

“The code execution sandbox normalized the amount of reasoning it takes to solve this particular problem, and so you can see a medium weight model which runs faster outperform a heavyweight model.”

“Tool adjustments can make these workflows run faster at higher quality on cheaper or more specialized models.”

“Just by configuring better tools, you'll get a faster and cheaper agent, which really matters at scale when you're running a ton of these sessions a day.”

[Akshay Krishnaswamy @hyperindexed]

“Getting a brief taste of ‘modern’ Windows feels like visiting Cuba or something.”

[John Gruber @gruber]

Apple’s “Pro and Tell” event in Los Angeles, the new MacBook Pros with PROLED XDR displays, and the state of the Macintosh platform.
https://t.co/xlF3xjudwJ

[Palantir @PalantirTech]

"Sovereignty means ownership over the value that your organization creates.”

At DevCon 7, Palantir launched the sovereign stack built to protect your alpha.

Fine-tuned models that beat frontier performance.

Compute infrastructure, agent stack, eval primitives — the full set of tools to control every layer of your AI operations.

Stop renting intelligence. Start owning it. Stay sovereign.

[Jawwwn @Jawwwn]

Palantir CEO Alex Karp on being anti-war:

"Our products are built to avoid war—meaning the adversary knows that they will have such pain inflicted upon them that it's not worth going to war, and also to create social cohesion because we're going to bring back our troops healthier, safer, non-damaged, and that makes our society much more effective."

"I grew up as a lefty. I don't actually like war. I want to avoid war. I don't support wars. Sometimes they are necessary. And when they are necessary, America and its allies need the very, very, very best tools."

"The way to stop a war is for them to know these things work."

"We do have a rival, an adversary on the battlefield. I happen to respect their culture quite a bit, China, but it is us or them. And they have a dynamic, interesting, super high IQ scaled culture that is doing exceedingly well at AI."

"We built products that would protect civil liberties and stop terrorism—for the very simple reason that you don't want to do one or the other."

"If you don't stop terrorism, you're going to get a radical fringe on the right. If you stop terrorism without rights, none of us want that."

[Palantir @PalantirTech]

Palantir reports Q3 ‘26 US comm revenue growth of 148% Y/Y and revenue growth of 96% Y/Y; raises FY ’26 revenue guidance to 91% Y/Y growth and US comm revenue guidance to 143% Y/Y, crushing consensus expectations.

US comm revenue accelerated to 29% growth Q/Q and was $984 million, up from $397 million a year ago and $179 million two years ago, and surpassed US government revenue for the first time ever.

We also generated $1.128 billion in Q3 ‘26 GAAP net income, an increase of 137% Y/Y.

[Chad Wahlquist @chadwahl]

Another report from MIZUHO

"A single shoe model remarkably involves ~8B design choices. Once PLTR's Ontology was in place, it produced a design in just 3 days. Changshin also emphasized that it, and not PLTR, owns its processes and data."

"Training with NVDA's Nemotron took 2 hours and raised performance from 83% to 86% at much lower cost... the resulting model could identify the fire, flag gas exhaust for aircraft to avoid, and identify the location of residents in need of evacuation."

"Using PLTR, it fine-tuned an LLM with the Ontology and switched to an open-weight model in 3 days."

[Palantir @PalantirTech]

Welcome, Startup Fellowship Cohort 003, and thank you to all who applied.

[Palantir @PalantirTech]

THE SOVEREIGN AI REVOLUTION IS IN FULL FORCE.

We are helping our customers build superintelligence they own, with data they control.

McCarthy Building fine-tuned Nemotron in Foundry on labeled RFI data in the Ontology to identify project risks before they escalate, outperforming frontier models at significantly lower cost.

Tampa General is using sovereign AI to better protect patient data, keeping sensitive data and model processing local, and fine-tuning Nemotron models to classify LLM traffic against its own risk and privacy standards.

NVIDIA is using Palantir AIP and Nemotron to accelerate its own supply chain, shortening the path from wafer to deployment, spotting disruptions before they hit production, and keeping its models under its control.

L3Harris is bringing AI workflows in-house and building its own models trained on proprietary data to support critical defense use cases and the warfighter.

Ondas is launching Regna AI on Palantir AIP, leveraging open-weight models including NVIDIA Nemotron, to help customers adapt, evaluate, and deploy mission-specialized AI while retaining control of their data, models, and infrastructure.

[Palantir @PalantirTech]

ROK pop-up shop no.2
hannam-dong, seoul
october 22 & 23

registration opens october 13, 10am kst

stay sovereign.

[Akshay Krishnaswamy @hyperindexed]

In an era of linguistically adroit machines, there’s something mildly endearing about boneheaded phishing / scam emails

[Ted Mabrey @MabreyTed]

When Palantir is the only company earning positive rewards in your RL environments

[Chad Wahlquist @chadwahl]

AI systems do not get better from benchmarks alone

You have to connect signal, interactions, intended outcomes, actual outcomes and build a continuous learning loop for your agents to get better over time

Build → run → trace → debug → manage → learn

You get these connected logs/data/process across business systems with the Ontology

Workflow Lineage + global log search make the loop visible for building, securing and maintaining

The same backplane captures the data to train and fine-tune your own models

[Akshay Krishnaswamy @hyperindexed]

Sovereignty

[Chad Wahlquist @chadwahl]

Counterexample from @nvidia's own supply chain:

They post-trained the smallest model, Nemotron 3.5 Lightning, and it outperformed Nemotron 3 Ultra on predictive correctness “by a large gap.”

Retraining on 1 GPU, a couple of hours.

Nemo + Palantir Ontology + Autopilot, running on-prem because the data is sensitive. A governed learning loop beats “just wait for the next general model” when the work is operational, private, and continuous.

[Akshay Krishnaswamy @hyperindexed]

@petewilz