🎉 Celebrating 100 users on TRAKKAR.IN! We're grateful to every team, freelancer, and startup that trusted us to improve productivity and simplify their workflow. and yes, it’s now FREE for 1 year! Let’s keep growing together 🚀 🌐 trakkar.in | 📞 91 8141067657 #TrakkarIN #MilestoneCelebration #100Users #TimeTrackingIndia #ProductivitySaaS #WorkSmarter #StartupIndia #RemoteTeamSolutions #MadeInIndia #GrowthTogether

This is big: all access to Mythos and Fable AI models disabled for everyone outside America. First thoughts: 1. Technology is the ultimate weapon. National sovereignty, national security, all of it is now about technology. 2. Globalization is dead and Bharat must find her own way ahead. We must keep these two ideas in mind. What can our government do right now? Ensure that orgs in India embrace smaller models, both Indian and Chinese open source ones. With a bit of effort, we can make them work. Anyway, why pay money to people who don't even want to sell to you? We must deepen our R&D. Sarvam has been on it and we have been on it but remember that the latest models cost not only huge GPU budgets to train, the GPUs themselves are restricted. So we can't afford the scale of money (of the order of $100 billion to even get in the game!) and even if we could come up with the money, we can't get all the GPUs. I would not like to ask the government to fund tens of billions of dollars on this anyway - the money has far better uses. Zoho has been pursuing alternative R&D approaches that are far, far less expensive but by its nature cutting edge R&D takes time and we are patient. I am confident we will get there. Any remaining people in India who have delusions about globalization should wake up now.

I am working on an awesome feature , you don't want to miss. May be it's time to change your project management tool. Stay tuned. Something big is coming soon! Demo video soon!

Practical implication: if you're building agentic systems, the model's internal signals at specific token positions may be a far better hallucination detector than asking "are you sure?" or reading log-probs. Paper: arxiv.org/abs/2604.22271

Replicated across Gemma 3 27B and Qwen 2.5 7B, on TriviaQA and MNLI. So this isn't a quirk of one model — it looks like a general property of how transformers organize self-evaluation.

The wildest part: causal interventions show the PANL signal can rescue error detection even when the answer info is corrupted. The signal isn't just "this is probably wrong" — it encodes whether the model has the knowledge to fix it.

They tested it with a verify-then-correct paradigm. Three results: 1. Verbal confidence predicts errors far better than log-probs 2. PANL activations predict errors better than verbal confidence itself 3. PANL predicts which errors the model can actually fix

Their finding: there's a specific token — the newline right after the answer (they call it PANL: post-answer newline) — where the model caches a confidence representation. This signal causally drives verbal confidence and dissociates from token log-probabilities.

The paper (Kumaran, Patraucean, Osindero, Velickovic, Daw — arXiv:2604.22271) borrows from decision neuroscience. Humans use a "second-order" model: a separate evaluative signal that can disagree with the committed response. Turns out LLMs do something similar.

The puzzle: LLMs sometimes detect and correct their own errors with no external feedback. But how? In a "first-order" system, confidence comes from the generation itself — so the chosen answer should always look maximally confident. Error detection shouldn't be possible.

A new paper from DeepMind researchers cracked open something fascinating: LLMs have an internal confidence signal that predicts whether they're wrong — and whether they can fix it — even when their own words say otherwise. Here's what the research found 🧵

Vibe coding is just paying a model to understand your codebase for you. Fine for a weekend prototype. Bad idea in production. The devs who actually read what the AI writes will outlast the ones who just paste it.

Trakkar.in update: 100 teams onboarded 0 paying (free for a year, my fault) Lesson: free users give feedback, paying users give signal Moving to paid plans next month. I'll share the churn numbers either way.

SaaS founders keep making the same mistake in 2026. Free tier, then "growth", then no revenue, then "we are pre PMF", then shut down. If 0 people pay you 5 dollars, 0 people will pay you 50. I made Trakkar.in free for a year and I regret it.

Most of the "AI agents that replace your team" on this timeline are one GPT call inside a nice looking UI. The real ones are quietly closing enterprise deals and never post about it. Pick which one you want to build.

820 posts. 49 followers. If you're building in public and the followers aren't coming, it's not the algorithm. It's that nobody knows what you're building or why they should care. Took me a year of Trakkar.in to figure that out.