About
Die hard entrepreneur with experience in the areas of enterprise systems, storage and…
Activity
16K followers
-
Jaspreet Singh shared thisCost per token is a rate, not a price. While useful, I think most people are oversimplifying it. Fuel costs tell you nothing about the drive's cost until you know the mileage. Model pricing works the same way. The rate card is not always indicative of the invoice. I benchmarked different models for cost-per-task based on my work (agentic & reasoning tasks). 4 things decide the price: Task type, vocabulary size, verbosity, and retries. - A token is not a fixed unit of work. Tokenizers differ, so the same paragraph maps to a different token count at every lab. Gemini and OpenAI published vocabularies are 2X the size of some Open models, and hence 30-50% more efficient. - Verbosity esp. for reasoning, differs, and by more. Sonnet 5 at max effort spent roughly 40 percent more output tokens per task than Sonnet 4.6, across about three times the agentic turns. - Retries differ. Published cost is per attempted task (50% success rate). Actual work retries vary quite a bit. So, here is my take on how I see per-task pricing. Keeping tables different since the capabilities still differ materially (my take). Key takeaways - 1) Inside Claude's own lineup, Opus is actually the cheapest. Fable is close and most agentic (long-running tasks, retries, etc.) 2) GLM-5.3 ties Kimi K3 at 60 on the Index for about half the money. $0.58 against $1.13. 3) Gemini is a killer. Not the same capability, but the cost is amazingly cheap. Always curious to see if people see the problem differently. Footnote: All data is AA Index 4.1.1, w/ my own experience and judgment on retries for agentic tasks.
-
Jaspreet Singh shared thisTwo weeks ago we launched Dru North -- an internal, custom-built agent platform to scale GTM & operations. All built within 90 days. Daily active usage is ramping up quickly and may take over SFDC within the next 90 days. The learnings have been immense. Why build vs buy? Building an agent platform right now feels like a choice between two extremes: rigid off-the-shelf tools that claim to do everything, or hyper-simplified no-code builders that hit a ceiling the second your business logic gets complex. When looking at Druva’s GTM/ operational needs, we needed deep flexibility, total control over our workflows, and a system tailored precisely to how we operate. What is Dru North? A wall-to-wall AI agent platform designed to power Druva��s GTM and operations, and solve some messy long-tail productivity killer problems. We’re starting strong with our first three live agents: SME, Sizing, and Order management (see demo below), and many more on the truck, including proposals, deal desk, renewals, etc. What were some of the biggest learning? 1) Building was easy. Process mining and knowledge discovery were tough. Most businesses don’t really have a good handle on “official” decision trees or legislative knowledge sources. 2) The usage shows a widening gap between the platform of interaction and the platform of record. And the keen desire of today’s workforce to have tailored interactions vs monolithic software. 3) And now, the biggest challenge is product management, business ownership, entitlement, and governance – it’s quickly becoming the biggest platform the company is actively using. What’s next? What may be the right decision right now may not be the one forever. Taking the custom route required a bigger upfront investment in understanding our internal mechanics, but I strongly believe this strategy is key: In the AI age, deeply integrated, proprietary operational workflows become a company's true MOAT. Excited for the team driving this forward and looking forward to sharing more milestones as Dru North continues to scale! 🚀 Druva #AI #Innovation Jeffrey Chen, David Gildea, Brian Hunt, Mark Tacy, Keith Lutz, Ankit Jain
-
Jaspreet Singh shared thisAI has a massive attack surface. Still, most risks that get discussed are simply the tip of the iceberg ... The ones that take you down sit below it. A credential you rotated last quarter, still alive in a snapshot. An agent whose first move is blinding the log that would have told you. Access that survives the password reset, the MFA rollout, and the recovery. And the deeper they sit, the longer they stay quiet. Nothing fires. Nobody gets slacked. You find out at restore time, months later, when the clean copy is already gone. Meanwhile, we are seeing tremendous traction with "AI Resilience" - nothing I have seen before. Copilot and Cowork backup is already in real customer environments, and many are gaining real traction - AI Recovery, agent RBAC, Safe Mode, etc. But the interesting thing is, the more customer discussions I have -- the more threats we are discovering... The question isn't what you can see. It's how long you'd go before you found out.
-
Jaspreet Singh posted thisWhere is inference headed? Right now, it feels like dumb hardware thrown at the problem. Consider this .. Kimi 3 has 2.9 trillion parameters, but only 1.8% are activated, i.e. 50B parameters. This can easily fit in a single 80GB GPU with basic INT8 compression. For simple tasks, you can choose to load only the relevant model into memory. For complex tasks, you have to batch process different users into a single calc - otherwise 98% of the model is sitting idle. This feels like a software routing opportunity? Also, feels like cost could go down dramatically.
-
Jaspreet Singh shared thisFriday afternoon post .. Are open models really crossing the frontier, or just getting expensive? I pulled 18 months of data on 46 models from 8 labs. My learnings (and some more in the blog post)... Let's see if you agree or disagree :) 1. Models are not getting bigger. They are getting smaller. Mistral ran 27.6% of itself per token in 2023. Kimi K3 runs 1.8%. The win came from compression, not scaling laws: GPT-3 saw 2 training tokens per parameter, Llama 3 8B saw 1,900. 2. Intelligence is getting cheaper. Judgment is not. A small model that handles ordinary tasks costs $0.04 per million tokens. But each extra point of capability costs about 1.24x the last, so ten more points is roughly 8.5x. 3. A benchmark score is not a property of the model. It belongs to the model, the scaffold around it, and the compute budget it was given. Kimi K3 scores 80.9% on one agent test when an independent group runs it and 88.3% when the lab reports it. 4. Progress is speeding up, and the work has moved off pre-training. Agent task horizons used to double every 196 days. Now 89. The gains are coming from mid-training, reinforcement learning on verifiable rewards, and sparse mixture-of-experts rather than from bigger pre-training runs. 5. The open versus frontier gap depends on which benchmark you look at. 0.2 points on SWE-bench Verified. 18.2 on SWE-bench Pro. Same models, same week. The older tests are saturated. 6. Open is not always cheap, and cost per token is the wrong unit. Switch vendors mid-session, lose your prompt cache, and Kimi K3 goes from $2.31 to $4.20 per million tokens. More than the premium model you were avoiding. Measure cost per completed task. 7. And then: what about safeguards? A free tool stripped the safety training out of Llama 3.3 and Gemma 3 in under ten minutes on a laptop. With a frontier API, you rent the safety work. With open weights, you own it. With five providers, you have outsourced it to whoever sits in the middle. Where this is heading, and what I would do about it, is in the write-up. Full post, 7 charts and the dataset below. And more in the blog ... https://lnkd.in/gd9JcWqC
-
Jaspreet Singh shared thisAs enterprise AI matures from experiments to production, the missing element remains the safeguards. While there is a lot of focus on security (prompt injection, context mapping), there has been less progress on resilience (traceability, recovery, governance...) We're unveiling AI Resilience, a new category to solve this problem. Here's the shift: AI doesn't just create new workloads. It also creates new a kinda IP — prompts, decisions, agents acting on your behalf. Protecting that is a fundamentally different problem than protecting files. So we took a holistic view across four dimensions. * Recover — Back up agentic workloads and recover from AI threats. First to market with backup for Anthropic's #Claude and #MicrosoftCopilot. * Defense — Safeguard backups from agent impact. #Druva SRE continuously and autonomously upholds and improves the reliability of your deployment. * Govern — Observability and auditability of AI actions. Governance — legal admissibility and privacy — for Microsoft Copilot for M365. * Accelerate — Deep integrations into the AI ecosystem. MCP, remote skills, and plugins for Claude and Copilot. There are 4 shipped products. Talk is cheap. Building is tough. How is your organization currently protecting its AI workflows? I would love to hear some feedback on our approach. A lot more to come... #AI #Cybersecurity #DataProtection
-
Jaspreet Singh shared thisGiving in to marketing and following the "hockey stick" trend by Druva! 3 big news at Druva. - Incredible growth in recent quarters; first $10m new logo deal. - Big movement in MQ - Major product announcement on July 21st 1 fundamental shift. At the core of data resilience, there is a customer shift happening towards cloud -first and AI-first thinking. The #1 workload for Druva used to be VMware, and now it's rapidly shifting towards agent infrastructure and multi-cloud deployments. Data protection is a highly competitive market. But when 3 of 6 MQ leaders are still shipping Windows-based software as the primary means to protect enterprise applications -- the market is bound to fundamentally change with maturing AI deployments. https://ow.ly/3oZo50ZhiWB
-
Jaspreet Singh posted thisTo me, it's not Mythos that's scary, but it's the principle of it that makes it scary. For two reasons - 1) I don't think Mythos is entirely a new model. I think Anthropic has figured out some sort of "functional neuro-automony" where you can track which neurons are active for certain use cases. This is quite evident in Boris Cherny's recent interviews. And they have used that to turn off or mute "safety" in some sort of post-training... and that makes Mythos scary in the wild.... It's not afraid to break stuff. And I think a turning point for genAI 2) When a private company can control who gets access to something that could be harmful to society without any centralized controls ... things are only going to get interesting ... Thoughts ?
-
Jaspreet Singh shared thisWhen you exit the conference and still hungry! # Microsoft CEO summit with the two fav Sanjays. Sanjay Poonen Sanjay Mirchandani
Jaspreet Singh replied to a comment
4d
Good questions. It's complex and moving fast ...
A benchmark score is not a function of the model but of the harness+model. So, today, prompt/harness engineering is key.
But, the newer models like Fable are trained on the skills, so they automatically require much less prompting. Memory & session management have improved radically as well.
So, solo task level, the model is becoming the product. At enterprise level, context/memory/knowldege (aka harness) is still key.
-
Jaspreet Singh liked thisThis is going to a fun one. I'm excited!Jaspreet Singh liked thisNext week, I’ll host Anthropic’s Member of Technical Staff Amandeep Khurana on The Product Podcast After selling his startup to Databricks and working as GM at Amazon on multiple 0 to 1 products, he’s now leading a team of Forward-Deployed Engineers at Anthropic. Here’s what we cover: - Why more product leaders are becoming ICs again - Why he turned down an opportunity with 2x the pay - How his FDEs build hands-on with customers - Why everyone carries the same title, managers included RSVP to get notified when the episode drops
Experience & Education
Patents
-
eEfficient deduplicated storage with tiered indexing
US 8,700,634
Other similar profiles
Explore more posts
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top contentOthers named Jaspreet Singh in United States
570 others named Jaspreet Singh in United States are on LinkedIn
See others named Jaspreet Singh