KyroDB

Recently Padraig Ó Catháin, Ronan Egan and folks published a paper presenting a survey on all the recent progress on Hadamard matrices. It covered my work also on the large number of Hadamard constructions and T matrices I had published. You can have a look. Although after its publication, I have also published the construction of 7 more open one's: 2572, 3356, 3436, 3820, 4204, 5912, and 6796. The speed at which field is moving: by the time you finish reading this, another one will be solved. arxiv.org/pdf/2610.27175v1

----+++-++-++-+--+-+-+-+-++-+--+--+---+++-+----+---+---++++++--+-+-++-------+-+---+----++-++-++-++-+-+-++-+-+++-----++--++-++++-++++++--++-++++++----+--++-++--++-+--+----+---+---++++++--+-+-++-------+-+---++++-++++++--++-++++++----+--++-++--++-+-+---+++-++----++-+-+-+++++++---+---+---+++---+++-++----++-+-+-+++++++---+---+---+++----++-++-++-++-+-+-++-+-+++-----++--++-+-+++--+---+++------+--+-+--+++++++-+-+-++----++-++-++-++-+-+-++-+-+++-----++--++-+---+++-++----++-+-+-+++++++---+---+---++++++-++++++--++-++++++----+--++-++--++-+-+-+++--+---+++------+--+-+--+++++++-+-+-++-+++--+---+++------+--+-+--+++++++-+-+-+-+----+---+---++++++--+-+-++-------+-+--- +---+++-++----+--+-+-+-+-++----+-------++-++---+---+---+-+++++-----+-----++--++-----+++-++-----++--+-++--+-++--+++++--+---++--+-+-+++---+-+-----++-+--+-++----+---+--++---+---+---+-+++++-----+-----++--++---+--+-+-+++---+-+-----++-+--+-++----+---+--++--+--+++++---+-+-++------+---+++-++++--++--+--+++++---+-+-++------+---+++-++++---+++-++-----++--+-++--+-++--+++++--+---+++-++--+--++++-++-+---++++-+++++--++-++-+--+++-++-----++--+-++--+-++--+++++--+---+-++--+--+++++---+-+-++------+---+++-++++-+--+-+-+++---+-+-----++-+--+-++----+---+-++-++--+--++++-++-+---++++-+++++--++-++-+++-++--+--++++-++-+---++++-+++++--++-++-+-++---+---+---+-+++++-----+-----++--++--- +---+-+-------++-+-+--++++++---+---+----+-+--++--+++++---+-+--+-+-+--+--+--+--+++++-+-++--++-----+----++++++-+---+++-++--++-+---++-+++---++++++-++-+-++-------+-+-+--+--++--+++++---+-+--+-+-+--+--+--+--++++-+---++-+++---++++++-++-+-++-------+-+-+--++-+-+---+---+-+++++--+-++----+++--++--+-++-+-+---+---+-+++++--+-++----+++--++--++-+-++--++-----+----++++++-+---+++-++--++--++---+---+++-+----+-+-+--++++-+++++-+--+-+-++--++-----+----++++++-+---+++-++--++-++-+-+---+---+-+++++--+-++----+++--++--+-+---++-+++---++++++-++-+-++-------+-+-+---++---+---+++-+----+-+-+--++++-+++++-+----++---+---+++-+----+-+-+--++++-+++++-+---+--++--+++++---+-+--+-+-+--+--+--+--++++ -+----+---+++-++++++--+-+-+++-----++-++--+-+-++--++-----+----++++++-+---+++-++--++-+--+-+---+---++++++--++++++---+---+-+--++---+++-++-++-+--+-+-+-+-++-+--+--+---++++-+-++--++-----+----++++++-+---+++-++--+++---+++-++-++-+--+-+-+-+-++-+--+--+---++++-++--++---++++--+-++-----+-+++-+++-+-+--+-++--++---++++--+-++-----+-+++-+++-+-+---+--+-+---+---++++++--++++++---+---+-+--+++-+---+--+--+-++-+---+-+--+-++----+-+----+--+-+---+---++++++--++++++---+---+-+--++-++--++---++++--+-++-----+-+++-+++-+-+--+---+++-++-++-+--+-+-+-+-++-+--+--+---+++++-+---+--+--+-++-+---+-+--+-++----+-+---++-+---+--+--+-++-+---+-+--+-++----+-+---+-+-++--++-----+----++++++-+---+++-++--++

There are too many false accusations and myths about data centres; let's debunk them one by one. - Water Usage: This whole water consumption thing is nothing but a hoax. Modern data centres use closed-loop direct-to-chip cooling combined with dry heat rejection; very early versions used evaporative cooling, which is very rare these days. In fact, a modern data centre requires water, which is equal to a medium-sized restaurant's daily consumption. - Energy and Power: "They consume too much power and drive up electricity costs for normal people". Of course, if you connect them to the regular electricity grid, which supplies power to homes, they will consume a large amount of power and drive up costs. The solution is to let these data centre companies or organisations build their own power plants. These data centres need gigawatt-scale power, so it is neither feasible nor sustainable for them to rely on the regular power grids, given the power required for them in the next 10 years. - Land Usage & Noise Pollution: There is enormous space available for a country of India's size for development of such structures. In fact, one can argue that data centres would fit in the highest ratio of land usage to job/value creation. The noise pollution thing is so silly; of course you can't be building these in front of residential areas. In rare cases where this may have happened, then the govt. need to bring laws to restrict those. Data centres are the foundation of AI progress, and AI is the greatest technology-building and wealth-building opportunity for humanity. Thus, India must not fall victim to these woke agendas, which the United States has already fallen into.

split the atoms, learn the world, transfer the mechanisms.

In Aug 2025, exactly 1 year ago, when I started @kyrodb, it all started with building a learned index, inspired by the paper of @JeffDean, 'The Case For Learned Indexes'. That index had performed 3x better compared to a B-tree; I just came online and shared those benchmark results; I had no idea that small tweet would get an appreciation from Jeff himself. Along with a few other factors, this appreciation was also a reason why I went full-time; the side project became a product after multiple iterations, and then a company. Thanks again, Jeff. The man inspires a whole league of researchers and engineers. Attaching that post in the comments.

Theorem 3.1 in Li, Kaba, Ravanbakhsh(AISTATS 2025) is FALSE. A smooth Gaussian counterexample gives two SCMs with exactly the same paired pre/post-intervention law, but incompatible required quotient graphs. Root-only perfect interventions do not identify the claimed causal abstraction. We also identify the failed equivariance step and give conditions under which a corrected identifiability result can hold. Paper in comments.

KyroBench paper is out, a benchmark focused on context correctness & safety-critical failures in real production agent/RAG workloads. Frontier systems still score 0 on certification. Link in comments.

Plain vector retrieval is not enough; memory systems help; graph systems help differently, but production context needs freshness, deletion, scope, authority, and proof.

DeepSeek V4 proved something significant with a 1M token context window. At 1M tokens, MRCR 8-needle accuracy drops to 0.59. That's a 41% failure rate on fact retrieval at depth. And that's after compressing the KV cache to 2% of the standard attention cost. So the needle problem at extreme depths still remains fundamentally unsolved by attention-based systems. V4's architecture is the best available evidence that the LLM itself cannot be the context system. Consider what CSA is doing: It compresses 4 tokens → 1 KV entry, ranks blocks by relevance, and drops the low-ranked ones. That is, in essence, a retrieval problem disguised as an attention problem. And DeepSeek solved it inside the model weights, meaning it's baked in, static, non-updatable, and blind to your actual knowledge freshness. But here's the catch: the model layer can compress context. It can retrieve better, but it cannot know that the pricing doc you fed it expired a week ago. The larger extrapolation of this is- as LLMs get bigger context windows, developers will stuff more into context, more docs, more history, more knowledge. The probability of stale/wrong information contaminating that context grows proportionally @kyrodb fixes this problem with a Context Runtime, giving AI agents fresh, verifiable, transaction-safe context as business knowledge changes Bigger context windows don't fix the accuracy/reliability problem; they make the surface area for stale facts larger.

🚨 78% of AI failures are invisible. That's right. AI gets something wrong, but no one catches it. Not the user, not traditional monitoring, not even a sentiment analysis. These failures cluster into recurring patterns: → The confidence trap- AI is confidently wrong, and the user accepts it → The drift- AI gradually answers a different question than what was asked → The silent mismatch- AI misunderstands but produces something plausible enough that the user doesn't push back These patterns persist across 93% of cases even with more powerful models, because they stem from interaction dynamics, how models present outputs, and how users communicate intent, not capability gaps. This is one of the core problems we are solving at @kyrodb . We believe, for the deployment of responsible and reliable AI, the real and critical bottleneck is the infrastructure around it, the harnesses with which the model interacts, not the model itself. We built the KyroDB runtime to fix one part of it- the WRONG CONTEXT. It contributes largely to the model hallucinating and even more blunder, answering something wrong that too, with confidence. KyroDB makes sure the context is fresh, safe, and reliable to use before it reaches your model. Check out KyroDB to stop your AI from lying with confidence. Link in comments.

Saw a reel in which a contestant pitched a context solution to @waitin4agi_ and he rejected that by saying 'Context won't be a problem in the long run' by giving an example of increasing context window in LLMs. I have great respect for Varun, but he is 100% wrong here. The idea of 'We can solve context problem by increasing the context window' is out of touch in so many ways. Bigger context windows help capacity, but they do not automatically solve selection, relevance, or pollution. 'Needle in the haystack and 'context rot' problems are one of the most complained issues when you visit developer forums(on X too). We are fixing this context issue at @kyrodb. Long context ≠ usable context Many frontier researchs also highlighted this issue. 🧵

Opening up my calendar to chat on AI agents, mostly around retrieval/context/memory. If you are someone working in this field or even interested in this, let's chat and discuss what's going on in the market. Link in comments.

@kyrodb is designed to solve exactly this problem by making retrieval adaptive, predictive, and cache-first instead of vector-first. It avoids that trap with Hybrid Semantic Cache + learned prediction + tiered coherence, which turns RAG into something that stays sharp even as scale grows, especially for the agentic/voice/real-time use cases where collapse hurts most.

Get safe cache + retrieval brain in one system for your agents. Get @kyrodb

Cache invalidation is supposed to be impossible for Vector databases; we solved it

HNSW + LRU is broken

High and precise recall is a necessity for agentic search now. For traditional search workloads (find me the top-10 similar movies), 97% recall is fine. For agentic reasoning loops where a single missed document can cause the agent to hallucinate or take a wrong logical step, even a 2% recall drop is potentially catastrophic. Every document the agent fails to retrieve is a reasoning failure that cascades through the entire chain. @kyrodb gives you exactly what these agents need, extreme high recall at high dimensions

Hey @bernhardsson, I had submitted the benchmarking PR on the 'ann-benchmarks' repo. Please accept the benchmarks. @kyrodb is a self learning vector database optimised for RAG and AI agents' workloads. It is performing SOTA across all datasets. Here- github.com/erikbern/ann-benc…