James Bond

Life is equal, treat others as you would like to be treated.
Duke, MBA

[James Bond @kunwarmahen1]

And this your response. You guys suck.
We'd suggest continuing discussion with the hotel directly. We've shared your feedback with them.
1

[James Bond @kunwarmahen1]

@Marriott first for me. I have a booking and no rooms available for 45 min and this is 11:30 in the night. This is a worst service I have ever had in my life . You guys suck.

[James Bond @kunwarmahen1]

@Marriott Sheraton new york times square hotel

[James Bond @kunwarmahen1]

@PeterSchiff He wanted to put tariff on USA and pocket the tariff amount.

[James Bond @kunwarmahen1]

@PankajjKashyapp At least some FDI (for funding protest). Good for INR

[James Bond @kunwarmahen1]

@sanjeevsanyal Sir, one of the main factors middle class is having 2 or less kids is cost of raising kid is very high and both the parents are working to support them. Until these costs come down, no incentive can stop the decline.

[James Bond @kunwarmahen1]

@AstroSharmistha @X Nikita left SpaceX few months back.

[James Bond @kunwarmahen1]

@sanjeevsanyal Rest 13% want’s anarchy.

[James Bond @kunwarmahen1]

@OfficialCustICE Fuck off you scammers. Be aware

[Aswath Damodaran @AswathDamodaran]

My annual country risk update, which includes everything I know and have learned about country risk in my corporate finance analysis and valuation is now available to download: https://t.co/w1ynvhc6t5

[James Bond @kunwarmahen1]

@ClientSpprtDesk Yes sucker , why don’t you d(i)e, trying to rip people.

[James Bond @kunwarmahen1]

@elonmusk my good friend had volunteered to go to Mars, can you please consider him. His name is Vijay, Vijay Deenath Chauhan from
Mandwa.

[Naina Singh @madhatter_ns]

Every infrastructure team uses percentiles for SLAs because average latency is a lie. But inference cost is reported as a single number.
The cost problem isn't just the average hiding its tail. It's that "inference" isn't one distribution. It's multiple workloads with completely different cost structures sharing infrastructure, sharing a budget line, and sharing a single cost metric that doesn't resemble them.
I've asked teams running millions of requests a day whether their latency SLA is the same across all their use cases. The answer is almost always a pause, not because they haven't thought about it, but because the question never surfaced.
The precision is the trap. A sharper average of the wrong category is still the wrong category.
Decompose inference into the workload shapes you actually run, and see the cost conversation change to where the money goes, where the margin is, what the next use case will actually cost. Numbers a CFO can act on instead of one that just survives the meeting.

[James Bond @kunwarmahen1]

@homam108

[Naina Singh @madhatter_ns]

Token prices fell 98% in two years. Enterprise AI bills tripled. Why? Because a simple lookup and a complex reasoning chain consume wildly different compute, and deliver wildly different value, but land on the same budget line.

Cost per token, revenue per head, tokens per user; they all measure volume. None of them answer "how much of what."

We won't manage the spend until we measure what the infrastructure accomplished per dollar.

https://t.co/Ok7wCEqIaP

[Naina Singh @madhatter_ns]

A code completion and a multi-turn reasoning chain consume orders of magnitude different compute and yet in most enterprise budgets, both show up as the same line item.
@ServiceNow burned through its full-year @AnthropicAI budget in months. @Uber did the same. Neither had user-level or team-level telemetry on the spend and the lack of telemetry is correctly identified as the first problem.
But even when that visibility arrives, it won't be enough. User-level telemetry tells you who is spending. Task-level telemetry tells you where the money goes. I've watched teams get the first and still not solve the second. Until we solve for task, we won't be able to build an effective system to control the spend.
https://t.co/68yyzaplxl

[James Bond @kunwarmahen1]

@vikramskr @madhatter_ns may be Naina the PM for llm-d can answer

[James Bond @kunwarmahen1]

@annamalai_k All the best and where we are there with you irrespective of what decision you take

[James Bond @kunwarmahen1]

@annamalai_k wish both us happy wala birthday. Looking forward to all the great things you will do and make 4th June borns proud.

[Naina Singh @madhatter_ns]

AI infrastructure gets tuned at launch for the workload you have at launch.

Take the batching window, how long the system waits to collect requests before processing them together. Wider means better GPU utilization, narrower means faster responses. Most teams get this right on day one.

Then the business changes. New teams onboard, new applications ship, new use cases arrive with different expectations. Nobody revisits it because nothing looks broken. Costs are stable, throughput is fine, and there's no signal telling anyone to go back and check.

But users on the newer workloads start noticing something is off. Responses aren't failing, they're just slower than they should be. Enough friction that people start working around it instead of through it.

By the time that shows up in your metrics, your users already felt it. You've felt that yourself, the gap between "something is off" and "the numbers caught up."

The settings that were right for the business you had at launch may not be right for the business you have today. Infrastructure that can't adapt to shifting workloads quietly becomes the bottleneck nobody is looking for.