Ashish

20 years in semiconductors. $10B+ impact. Now building engineering-grade AI where proof matters.
#TruthIsAllYouNeed

[Ashish @ashish_iitk]

@KobeissiLetter The digital employee idea only works if the agent can be trusted around real company systems. Once it touches data, tools, and workflows, the hard part becomes access, context, ownership, and knowing exactly what it did.

[Ashish @ashish_iitk]

@aakashgupta This is a very real shift. As models get better, the scarce thing is expert judgment on whether the work is actually right. AI can produce the answer, but someone with domain depth still has to know if it makes sense.

[Ashish @ashish_iitk]

@GaryMarcus This is a good reminder that “same question, different language” is not always the same test.
The model is pulling from different patterns, sources, and cultural defaults depending on how the question is framed.

[Ashish @ashish_iitk]

@levie This matches what I’m hearing too. The hard part is not getting agents into the company, it’s figuring out who owns them, what data they can use, and how you know they’re creating real value.

[Ashish @ashish_iitk]

@dee_bosa This is where enterprise AI gets real. Public benchmarks are useful, but every company has its own codebase, workflows, risk, and edge cases. The best model on paper may not be the best model for your actual work.

[Ashish @ashish_iitk]

@PolicyEngineUS This is a useful benchmark because taxes and benefits are exactly the kind of work where close enough is not enough. Small errors can change real money for real people.

[Ashish @ashish_iitk]

@aaditsh Open models made sense while China was catching up, but once the gap gets smaller, access becomes the thing everyone starts protecting

[Ashish @ashish_iitk]

@gkcs_ Yeah, this is the hard part. AI is useful, but the spending is so massive that useful alone may not be enough. The returns have to be huge too.

[Ashish @ashish_iitk]

@rohanpaul_ai AI usage is not the same as AI value. A lot of companies are now learning that more tokens used is not a productivity metric.

[Ashish @ashish_iitk]

@OurWorldInData @redouad AI model choice is becoming less about brand loyalty and more about availability, cost, performance, and control.

If Chinese open models keep gaining usage, the default model assumption changes fast.

[Ashish @ashish_iitk]

@edzitron If this mainly applies to logged-out traffic, that changes the story a lot. Cutting cost on low-value usage is useful, but it doesn’t prove the economics of heavy paid or enterprise workloads.

[Ashish @ashish_iitk]

@alliekmiller Prompt engineer was always too narrow.

The real role is closer to someone who understands the work, the tools, the risks, and can turn AI from a toy into an operating workflow.

[Ashish @ashish_iitk]

This is the version of AI adoption people don’t talk about enough.

The best companies don’t just replace work with AI. They use AI to increase capacity, move faster, and create more work worth staffing around.

[David Sacks @DavidSacks]

Narrative violation: A new study of 21,559 firms in the U.S. finds that “companies that adopt AI tend to grow faster following adoption”.

“Firms making the largest AI investments grow employment by roughly 10% following adoption, while low-intensity adopters see no statistically significant change.”

“Entry-level headcount rises 12% for high-intensity adopters.”

“Gains emerge gradually and are broad across roles, including engineering, sales, administration, and customer service.”

“The results counter predictions that AI adoption will lead to broad job loss.”

The study is based on observed AI spending from Ramp card and bill pay data linked to Revelio Labs workforce records.

[Ashish @ashish_iitk]

@kareem_carr This is exactly right. AI can do more tasks now, but responsibility does not transfer with the task. Someone still has to know whether the work is correct, safe, and actually done.

[Ashish @ashish_iitk]

AI can be transformational and still have parts of the investment cycle overextended. The real question is how much of this capex turns into durable revenue before the market runs out of patience.

[Global Markets Investor @GlobalMktObserv]

🔴The AI investment boom is unprecedented, but questions over its sustainability are growing:

At the current pace, AI investment has already surpassed every prior technological boom in history at the same stage, rising to more than 4.5 times its pre-boom level in just 3 years, according to BIS data.

By comparison, canal mania in the 1830s and railway mania in the 1840s peaked at ~3-4 times their pre-boom levels, and both ended in economy-wide recessions.

This comes as the 5 largest hyperscalers are expected to invest more than $1 TRILLION combined from 2025 through the end of 2026.

Meanwhile, the Bank for International Settlements warns that if returns disappoint, financing could be pulled back rapidly, turning the capital expenditure boom into a prolonged bust that could rattle financial markets and damage the global economy.

Furthermore, a major AI-related equity correction could have WAY broader consequences than in prior cycles, since households now hold significantly more of their wealth in stocks than in previous decades.

History does not repeat, but it often rhymes.

[Ashish @ashish_iitk]

@oguzerkan This is the better way to look at it. AI capex can be overextended and still be attached to real revenue growth.

The mistake is treating this like either pure bubble or guaranteed payoff. It is probably both: real demand, with some bad assumptions around it.

[Ashish @ashish_iitk]

This is the part every AI business has to face. If intelligence gets good enough and cheap enough, the moat can’t just be we have a better model.

The moat moves to distribution, trust, workflow ownership, and what the system can actually do for the user.

[Xiaoyin Qu @quxiaoyin]

China’s AI playbook: kill OpenAI and anthropic with free great models. Make it free. Then use cheap electricity to export compute as well. Currently the blocker is chip but Hauwei would catch up soon. Imagine a world where instead of paying hundreds of billions to OpenAI and anthropic, you pay almost zero to similar level of intelligence with cheap cheap inference. What’s gonna happen?

[Ashish @ashish_iitk]

@GaryMarcus If models keep converging this fast, the moat can’t just be the model. It moves to the layer that turns models into reliable workflows people actually depend on.

[Ashish @ashish_iitk]

@kmeanskaran This is the part people miss. AI can help build the system, but it cannot replace knowing what system should exist in the first place.

That still needs judgment, domain context, and someone accountable for the outcome.

[Ashish @ashish_iitk]

@rohanpaul_ai AI did not fail here because inspection is impossible. It failed because manufacturing has years of tacit knowledge hidden in edge cases, supplier mistakes, and “I’ve seen this before” judgment.

That knowledge is hard to automate until you know how to capture it.