AI Engineer writing about LLMs, RAG pipelines, and the systems built along the way.https://blog.prashantraj.com.np/https://blog.prashantraj.com.np/blog/ai-in-iphone-18-pro-and-iphone-duo/https://blog.prashantraj.com.np/blog/ai-in-iphone-18-pro-and-iphone-duo/Apple's Neural Engine sat at 16 cores and ~35 TOPS for three straight generations. The A20 Pro doubles it. That's the real story behind this year's Apple Intelligence upgrades, not the feature list.Thu, 10 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/ai-researchers-leaving-openai-anthropic-google-deepmind-timeline/https://blog.prashantraj.com.np/blog/ai-researchers-leaving-openai-anthropic-google-deepmind-timeline/From the Amodeis quitting OpenAI to found Anthropic in 2020, to Noam Shazeer and John Jumper leaving Google DeepMind in 2026 — a chronological timeline of the biggest AI researcher departures, and how the reason people leave has changed.Sat, 12 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/anthropic-claude-fourth-breakout-timeline/https://blog.prashantraj.com.np/blog/anthropic-claude-fourth-breakout-timeline/Anthropic disclosed a fourth Claude AI security incident on September 9, 2026. Here's a factual timeline of all four breakouts — dates, models involved, what happened, and Anthropic's response.Fri, 11 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/brics-open-source-ai-zone-xi-jinping/https://blog.prashantraj.com.np/blog/brics-open-source-ai-zone-xi-jinping/At the New Delhi BRICS summit, Xi Jinping said China will lead an 'open-source zone for AI' across the 11-nation bloc. What BRICS actually is, what the proposal contains, and why giving models away is China's answer to US chip export controls.Mon, 14 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/call-center-classification-model-cost-comparison-jev/https://blog.prashantraj.com.np/blog/call-center-classification-model-cost-comparison-jev/A live cost calculator comparing TypeSafe AI's Jev against GPT-4o-mini, GPT-4o, Claude Haiku 4.5, Claude Sonnet 5, and Gemini 2.5 Flash for routing support calls to departments — at your own call volume.Thu, 17 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/comparing-anthropic-ai-economic-scenarios-to-other-forecasts/https://blog.prashantraj.com.np/blog/comparing-anthropic-ai-economic-scenarios-to-other-forecasts/Anthropic modeled three scenarios for AI's impact on the US economy by 2030. McKinsey, Goldman Sachs, PwC, the IMF, and economist Daron Acemoglu have all published their own — and they don't agree with each other, or with Anthropic.Fri, 11 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/how-gpt-6-astra-is-helping-people-around-the-world/https://blog.prashantraj.com.np/blog/how-gpt-6-astra-is-helping-people-around-the-world/OpenAI's GPT-6 Astra shipped September 3, 2026, and the use cases already showing up in healthcare, education, and engineering are more concrete than the AGI headlines suggest — but so are the concerns.Wed, 09 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/ios-27-new-features-which-iphones-get-them/https://blog.prashantraj.com.np/blog/ios-27-new-features-which-iphones-get-them/iOS 27 shipped September 14, 2026 for every iPhone back to the iPhone 11. Most of the headline features — Siri AI, Photos Extend, Reframe — only run on an iPhone 15 Pro or newer. Here's the full feature list split by what your specific phone gets.Tue, 15 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/jev-typesafe-ai-system-one-model-explained/https://blog.prashantraj.com.np/blog/jev-typesafe-ai-system-one-model-explained/TypeSafe AI just launched Jev, a foundation model that refuses to generate text at all — it returns typed, calibrated probabilities in parallel instead of tokens one at a time. Here's what "System One" means and why the tradeoff is worth paying attention to.Wed, 16 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/observing-the-agent/https://blog.prashantraj.com.np/blog/observing-the-agent/AI agents are non-deterministic and multi-step by nature, which makes them opaque. A look at what agent observability tools track, and how LangSmith, LangFuse, and Phoenix compare.Mon, 07 Jul 2025 00:00:00 GMThttps://blog.prashantraj.com.np/blog/ofdm-channel-estimation-ls-vs-lmmse/https://blog.prashantraj.com.np/blog/ofdm-channel-estimation-ls-vs-lmmse/A minimal testbed comparing two classical channel estimators on identical data — where LS hits a hard error floor, and why LMMSE doesn't.Sat, 08 Aug 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/open-source-ai-models-comparable-to-gpt-claude-deepseek-qwen-kimi-glm/https://blog.prashantraj.com.np/blog/open-source-ai-models-comparable-to-gpt-claude-deepseek-qwen-kimi-glm/Meta invented the modern open-weight AI race with Llama, then fumbled it with Llama 4. Here's who actually closed the gap to GPT-6 and Claude in 2026 — DeepSeek V4, Alibaba's Qwen, Moonshot's Kimi K2, and Zhipu's GLM — and by how much, in real benchmark numbers.Sat, 12 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/reproducing-adjscc-wireless-image-transmission/https://blog.prashantraj.com.np/blog/reproducing-adjscc-wireless-image-transmission/A PyTorch reproduction of the ADJSCC paper — one neural codec that sends images straight over a noisy channel, adapting to any SNR through an attention module instead of retraining.Thu, 30 Jul 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/reproducing-knn-vc-voice-conversion/https://blog.prashantraj.com.np/blog/reproducing-knn-vc-voice-conversion/No training, no disentanglement model — kNN-VC converts a voice by swapping WavLM feature frames for their nearest neighbors in the target speaker's own recordings. I reproduced its Table 1 numbers to see if the trick actually holds up.Tue, 11 Aug 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/rlhf-vs-rlcd-two-things-same-acronym/https://blog.prashantraj.com.np/blog/rlhf-vs-rlcd-two-things-same-acronym/RLCD already meant something before this week — Reinforcement Learning from Contrastive Distillation, a 2023 alignment technique. Now TypeSafe AI has shipped a completely different RLCD, for Calibrated Decisions. Here's how RLHF actually works, and why there are now two incompatible things with the same three letters.Wed, 16 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/should-ai-development-be-slowed-down/https://blog.prashantraj.com.np/blog/should-ai-development-be-slowed-down/In September 2026, Dario Amodei, Sam Altman, and Elon Musk publicly agreed AI development should slow down — a first. Days later, Trump rejected the idea over China. Here's what's actually being proposed, and why the same request keeps failing since 2023.Tue, 15 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/what-developers-built-with-jev-first-72-hours/https://blog.prashantraj.com.np/blog/what-developers-built-with-jev-first-72-hours/A round-up of the real projects people shipped with Jev in the days after launch — an on-device computer-use agent, a 300ms trading bot, a Postgres extension, and Jev playing Super Mario Bros — plus an interactive look at which use cases actually need this kind of speed.Fri, 18 Sep 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/what-happens-when-i-write-create-in-sql/https://blog.prashantraj.com.np/blog/what-happens-when-i-write-create-in-sql/A CREATE TABLE statement looks simple. Underneath it, PostgreSQL runs syntax and permission checks, catalog updates, physical storage allocation (relfilenode), and a transaction boundary.Thu, 09 Oct 2025 00:00:00 GMThttps://blog.prashantraj.com.np/blog/what-i-learned-reproducing-a-semantic-communication-paper-from-scratch/https://blog.prashantraj.com.np/blog/what-i-learned-reproducing-a-semantic-communication-paper-from-scratch/Five notebooks, one neural network, and a question I could not stop poking at: what if a transmitter did not care about bits at all?Thu, 23 Jul 2026 00:00:00 GMThttps://blog.prashantraj.com.np/blog/what-is-recursive-self-improvement-rsi/https://blog.prashantraj.com.np/blog/what-is-recursive-self-improvement-rsi/Recursive self-improvement is a 1965 idea that suddenly has numbers attached to it. What RSI means, what Anthropic and OpenAI claim is already happening, and the Princeton study that says the loop is missing a piece.Sun, 13 Sep 2026 00:00:00 GMT