https://jottings.vishalvshekkar.com/ 2026-09-04T18:33:47.897Z https://jottings.vishalvshekkar.com/jots/5722626097.html 2026-09-04T18:33:46.260Z

Balancing cache writes-vs-utilization, tool calling back-and-forth, reasoning effort, to improve subjective & perceived sense of performance of an LLM-backed agentic workflow, while having a classifier to route each inference call to a complex hierarchy & classes of LLMs, providers, and wrappers which also reroutes during outages, degraded outputs, and latencies, while ensuring a consistent personality to the end consuming entity is the funnest engineering challenge today.

https://jottings.vishalvshekkar.com/jots/7788510791.html 2026-08-25T14:18:05.107Z

Simplicity is seldom achieved through simple design. There is quite a huge distinction between naive and earned simplicity. Naive simplicity comes from not engaging with the complexity. While simplicity is earned from understanding the complexity deeply enough to remove, absorb, or hide it.

https://jottings.vishalvshekkar.com/jots/3020589766.html 2026-08-22T17:30:05.897Z

My old memory foam pillow had started losing its form. It’s developed amnesia, if you will.

https://jottings.vishalvshekkar.com/jots/7850742765.html 2026-08-16T08:15:07.427Z

This compelling paper argues how the AI-automation arms race that organizations are having would lead to a depletion of spending power of the population and eventually collapse of the economy. Most people who are running these organizations know this eventual reality, but they remain trapped in an automation arms race to secure immediate cost savings for as long as possible, otherwise they would lose to their competition. That prevents them from exercising voluntary restraint. They propose a **Pigouvian automation tax** set equal to the uninternalized demand loss per task. Here, the *'uninternalized'* means the loss in consumer demand due to income loss of workers impacted by the collective industry decision of moving towards automation. The burden is shared by all the organizations involved in automating that task category. This, from what I understand, is calculated per task category where automation efforts are underway. The revenue from this tax, they propose, can be recycled to fund targeted worker retraining programs, which increases the economy's income-reabsorption rate and makes the corrective tax potentially self-limiting and transitional over time. **The AI Layoff Trap** by *Brett Hemenway Falk* & *Gerry Tsoukalas*

https://jottings.vishalvshekkar.com/jots/6701266184.html 2026-08-16T05:03:32.661Z

All that being said, the math behind watermarking is quite remarkable. Using the entropy in human language to encode some "unseen" information to communicate something is quite novel. The artifact linked below is by Thariq from the Claude team. It illustrates how watermarking works and how reliable it is for each type of generative model use case. https://claude.ai/code/artifact/803916fd-3bc1-465f-8738-d4ece6fc5071

https://jottings.vishalvshekkar.com/jots/6426488348.html 2026-08-16T04:17:44.883Z

There are watermarking approaches developed specifically with the objective of user attribution. [PersonaMark](https://arxiv.org/abs/2409.09739) and [Watermarking Language Models for Many Adaptive Users](https://arxiv.org/abs/2405.11109) are a couple of them employing interesting mechanisms to achieve that objective.

https://jottings.vishalvshekkar.com/jots/6365765748.html 2026-08-16T04:07:37.657Z

Anthropic uses SythID-Text approach for watermarking popularized by Google Deepmind, and tested within Gemini. This method uses tournament sampling to pick the next token from a possible set of tokens (size dependent on available entropy for that next token.) SynthID like many other watermarking technique uses a pseudorandom key to run the tournament knockout on the highly-likely candidates for the next token. In principle, generative text watermarking can be extended from just provenance detection ('*this came from model E*') to cryptographic fingerprinting ('*this came from account AGHRS125*') or even multi-bit attribution ('*this encodes identifier Y*'). Model providers may use different keys per account instead of one key across all generations. Possibly using a pseudo-random function that mixes the master key with a user identifier so they may generate it deterministically during detection/attribution. Another, more novel way is to encode some identifying information as binary payload into the watermark itself. `watermark payload = providerID | modelID | accountID | requestID` Just the way the plain watermark is encoded in the statistical manner determined by SynthID in the next-token available entropy space, that watermark itself could carry `accountID` and a lot more info. This way, any meaningful length of text generated by a model could have the signature of who prompted its generation. This is an extremely powerful tool, which I'm sure, would be exploited to its depths in the near future. Note: The SynthID paper itself doesn't get into personal attribution. But, there are other mechanisms built specifically with the objective of identification.

https://jottings.vishalvshekkar.com/jots/6254443759.html 2026-08-16T03:49:04.437Z

While I understand why watermarking is important to identify content originating from a generative model and establish provenance, it is yet another tool in that may be deployed in mass surveillance with ability to tie generated content to specific account/groups/geography. As much as I hate AI slop, and the deeply disturbing side of generative AI like deepfakes and mass propaganda it can enable, a tool to fight that can also be used as a tool for also targeting individuals.

https://jottings.vishalvshekkar.com/jots/1853780422.html 2026-08-15T15:35:37.804Z

Another powerful graphic journal picked up at the Bulbul: **Palestine by Joe Sacco**

https://jottings.vishalvshekkar.com/jots/1825539375.html 2026-08-15T15:30:55.393Z

Today's picks at Bulbul. The Blue Book, An Anti-Caste Graphic Novel by Siddhesh Gautam.

https://jottings.vishalvshekkar.com/jots/1780238722.html 2026-08-15T15:23:22.387Z

[Bulbul Books & Curious](https://www.instagram.com/thebulbulbookstore/) is probably my favourite bookstore. A big part of that is, of course, that it’s curated by my good friends Aishwarya and Abhishek. But more than that, they’re both extraordinarily well-read, thoughtful, introspective people, and you can feel that in what makes it onto the shelves. The curation is eclectic and genuinely curious. Every visit feels like a chance to stumble upon something I wouldn’t have gone looking for myself. It’s also a small bookstore, and I love it for that. There’s enough to discover, but never so much that it becomes overwhelming. You can take in the shelves slowly, follow your curiosity, and actually enjoy the pleasure of finding something unexpected.

https://jottings.vishalvshekkar.com/jots/1273926195.html 2026-08-15T13:58:59.261Z

Working on something in public! I call it Open Reverie. It's a place for your reveries. A companion agent harness with hierarchical tiered memory so it remembers a lot of what you've shared, and helps you reflect on it, come back to things, and build a graph of your thoughts. This was thought of as an extension to [Ruminate](https://ruminate.life) a few months ago, and as the librarian in [Murmur](https://murmur.you), but now it's also an open source harness for anyone to use. Would love code & idea contributions from people!

https://jottings.vishalvshekkar.com/jots/1769623954.html 2026-08-13T07:48:16.239Z

LLM watermarking turns the model’s linguistic entropy into bandwidth. Wherever the model has multiple plausible ways to say the same thing, those degrees of freedom become a channel through which a hidden (watermarking) signal can be transmitted.

https://jottings.vishalvshekkar.com/jots/1759735754.html 2026-08-13T07:46:37.357Z

Advanced LLM watermarking is closer to spread-spectrum radio than to stamping text with a signature. The language itself is the carrier; a weak cryptographic signal is spread across thousands of otherwise natural token choices. To everyone listening normally, it’s just language. With the right key, you can correlate against the hidden signal and pull it out of the noise. The text is not where the watermark sits, it is the carrier wave. The watermark lives in tiny coordinated deviations in how that carrier is generated. Without the key they look like ordinary sampling noise. With the key, they line up coherently into a signal.

https://jottings.vishalvshekkar.com/jots/1560720120.html 2026-08-13T07:13:27.201Z

Watermarking an LLM's output, in [Kirchenbauer et al.](https://arxiv.org/abs/2307.15593), is basically turning token generation into a dynamic casino where the house secretly nudges the model toward green tokens. A z-test detector regenerates those context-dependent green sets and asks whether the generated text landed on them far more often than chance would plausibly explain. It isn't detecting “AI style”; it's detecting an intentionally planted statistical bias.

https://jottings.vishalvshekkar.com/jots/4782401439.html 2026-06-22T14:10:24.014Z

Best way to deal with needing to separate clothes by colour for laundry… have an all-dark wardrobe.

https://jottings.vishalvshekkar.com/jots/6499276369.html 2026-05-29T11:36:32.763Z

A `isFavorite: Bool` is the most violent compression anyone's ever done to a feeling. "I loved that shot of espresso on a rainy Tuesday when the grind finally clicked" gets crushed down to a single byte. The heart icon is a one-byte shrine to a memory it can't actually hold. The sad, nerdy part is that we know one byte is nowhere near enough to be true, and we also know it's exactly enough to be useful. So we ship it. A real model of why you loved something can't be computed, and even if it could, nobody's tapping through twelve screens to log it. A faithful record of the feeling is useless. A sliver of a truth you can query is gold. That's what a favorite button really is. A quiet admission that we'd rather keep a searchable lie than an unsearchable truth. And that's most of data modeling, honestly. The question is never "how do we represent reality?" It's "how much of reality can we throw away before the query stops coming back?" > PS: A Bool is represented using a byte in most languages as opposed to a bit because the CPU can't directly address a bit.

https://jottings.vishalvshekkar.com/jots/5068007929.html 2026-05-29T07:38:00.079Z

*Amdahl's Law* says the speedup you get from throwing more workers at a problem is capped by the part that can't be split. You can run a thousand agents in parallel and still sit there waiting on the one commit that has to happen in order. It's the same as a company hiring a thousand engineers and still grinding to a halt because one person/team understands the deploy script. Parallelism doesn't kill the bottleneck. It just clears everything else out of the way until the bottleneck is the only thing left standing, lit up, impossible to ignore. The more you speed up the easy parts, the more obvious it gets that the slow part was always the drag. Most of what we call "scaling" is just elaborate work to avoid looking straight at the one step we never wanted to deal with. We'll automate, distribute, and optimize around it forever, because that feels like progress, and staring at the actual problem doesn't. Solving and optimizing for that is the real efficiency-unlock.

https://jottings.vishalvshekkar.com/jots/4820179286.html 2026-05-29T06:56:41.792Z

Philosophers spent two thousand years on the Ship of Theseus: *replace every plank one by one, is it still the same ship?* A programmer heard the question, ran the planks through a hash function, and said "*no, it's a different ship, here's the receipt, can we go to lunch.*" Turns out the riddle was only hard because everyone insisted the answer had to be profound. Ask "the same in what sense, and who's asking" and it stops being philosophy and becomes inventory.

https://jottings.vishalvshekkar.com/jots/2435498606.html 2026-05-24T09:12:34.986Z

The latest washing machines and dryers have AI sensing, moisture detection with material type care, heat pumps for energy efficiency, and many other advanced features. All I want (have wanted since I was a kid) is to have clear peek-in doors, illumination, and/or a camera that shows what’s happening inside. That’s the simplest thing they could add, and boom! It’s now an entertainment/meditation appliance, not just utilitarian. Same applies to dish washers.