Shival Tailor

tinkerer | views are of my own

[Shival Tailor @shival]

The best leaders are not the fastest talkers in the room. They notice when a quiet person stops contributing, and they ask why before the silence becomes a resignation. Retention starts long before an exit interview.

[Shival Tailor @shival]

AI is now producing results faster than experts can verify them. That flips the scarce skill. Generating an answer is getting cheap. Knowing which answers deserve trust, and building the checks around them, is where authority will move.

[Shival Tailor @shival]

On Oct 15, Commerce agreed to reallocate up to $955M in CHIPS incentives to GlobalFoundries, with GF raising its Malta, NY and Burlington, VT investment to $5B. The total award did not grow. The bet on photonics and packaging got more specific.

[Shival Tailor @shival]

NVIDIA RTX Spark laptops went on sale globally today, Oct 16, from ASUS, Dell, HP, Lenovo, Microsoft and MSI, starting at $2,599. Up to 128GB unified memory and 1 petaflop of FP4 in a laptop. Local agents just got real hardware.

[Shival Tailor @shival]

Meta released Muse Spark Open on Oct 15: 816B total parameters, 42B active per token, 1M token context, live day one on seven inference providers. The launch is not just a model. It is a distribution move against DeepSeek, Qwen and GLM.

[Shival Tailor @shival]

Good managers do not remove uncertainty. They name it early enough that the team can work with it instead of pretending around it. Most bad weeks start with a risk everyone saw and no one said out loud.

[Shival Tailor @shival]

Google is crediting Gemini 4 Argon with helping refactor roughly 6 million lines of Chrome code in weeks, not months. The shift is not just coding speed. Legacy code, once a liability, is becoming the proof point for what these models can actually do.

[Shival Tailor @shival]

AtScale launched AtScale for AI today with a blunt point: a model cannot reason its way to your definition of revenue. Enterprises do not just need smarter models. They need shared definitions, governed calculation, and a way to test accuracy.

[Shival Tailor @shival]

Colt DCS and RMZ started construction today on Mumbai 3, a 64MW AI-ready data center in Navi Mumbai. It is already fully pre-let. Under their $1.7B joint venture they plan 250MW across Mumbai and Chennai. Demand is being built before it arrives.

[Shival Tailor @shival]

Cohere said today it signed a letter of intent to buy AISLE, whose AI found more than 500 bugs in widely used software in its first year, including 12 in OpenSSL. Security is becoming a core model capability, not a feature bolted on later.

[Shival Tailor @shival]

The hardest part of leading is not making the decision. It is living with the second-order effects in public while everyone else judges only the first one. Good leaders leave room to revise without pretending they never chose.

[Shival Tailor @shival]

Two launches today point the same way. DeepMind puts scam detection on the phone, and BrainChip's Sequora runs models up to 3B parameters on battery-powered devices. The cloud won on scale. The edge will win on trust, latency, and privacy.

[Shival Tailor @shival]

On Tuesday, AMD became the financing backstop for Fermi's Project Matador. Its credit support covers the 222 MW first phase leased to TensorWave, with a framework to expand to 650 MW. The catch: AMD's obligations start only once construction finishes and rent begins.

[Shival Tailor @shival]

ICON raised $200M in Series D funding co-led by Baillie Gifford and BlackRock-managed funds to expand robotic construction. It says it has completed over 270 3D-printed buildings. Capital is moving from AI that writes to AI that builds.

[Shival Tailor @shival]

Google DeepMind and Singapore's GovTech is building an on-device scam detection model designed to flag manipulative chats, built on Gemma. It works without uploading private messages. The most useful AI this week may be the one that never leaves your phone.

[Shival Tailor @shival]

Worth an hour: Ruflo, an open-source orchestration layer for Claude Code and Codex, got a major release today. It coordinates fleets of agents with shared memory, and its new Mods turn one mission into a research, build, test and secure run. MIT licensed.

[Shival Tailor @shival]

Two frontier launches in two days, and both were pricing events. Google put a deadline on its intro price. Meta kept its best model off the API entirely, inside a subscription. Capability is becoming table stakes. Distribution is the strategy.

[Shival Tailor @shival]

IDC estimates enterprise AI spending reaches roughly $400B this year. McKinsey's 2025 survey: 88% of companies use AI somewhere, 62% are experimenting with agents, only 23% have scaled agents in one function. The bottleneck is no longer the model. It is operations.

[Shival Tailor @shival]

Rocket Software is acquiring Geniez AI: six people, one year old, a deal reported at up to $100M. Geniez builds agents that run mainframes. The scarce asset in enterprise AI is not models. It is people who understand the systems everything still runs on.

[Shival Tailor @shival]

Google launched Gemini 4 Argon today at $2/$10 per million tokens. The catch is the calendar: that price doubles after November 30. Every benchmark posted this week is being quoted at a price that expires in seven weeks.