
Finding the cadence of life. Hoarder of books, stories and experiences; entrepreneur, AI across multiple domains. @UF alum; Building @latentsig
Joined January 2016
- Tweets29.3K
- Following6.7K
- Followers884
- Likes53.9K
huduga retweeted
What the heck is happening?? 🚨
Gemini 4 Neon > Fable 5.2
Gemini 4 Neon > Opus 5.5
Gemini 4 Neon > GPT-6.2 Astra
wtf @grok , is it really true?
huduga retweeted
What a malicious campaign this is! Sarvam has clarified that its govt subsidized GPUs continue to be for sovereign model building and that the Sarvam token factory is different and comes from the series funding they got from VCs and commercial players.
Replying to @cneuralnetwork
You can hate me & I’m not doing a witch hunt - the employees call me names I’m just highlighting
Plus the 4K GPUs are part of government incentive to build sovereign ai so every citizen has a right to question them there
Yc funding is private and startups do pivot and it’sok
huduga retweeted
I'm sorry but Gemini 4 Neon is not on Opus 5.5 level.
It's above it, and it's cheaper.
After deeply testing Gemini 4 Neon for the past few days I can confidently say that this model beats GPT-6.2 Astra and Claude Opus 5.5 with ease.
Here is what I did with it in the last 3 days:
- Built two native iOS apps one-shot (zero bugs, custom player, Live Activities, widgets, Liquid Glass)
- Audited all my coding projects and it found so many bugs and performance optimizations, always choosing the best libraries
- My router has lower latency and faster upload now, and my Hermes Agent runs noticeably smoother
- On my Surface it analyzed the Win11 26H1 VadS kernel memory leak and fixed it with KB5132059
It talks better, does awesome research, explains everything clearly and has great world knowledge. I am not concerned at all about Anthropic, but OpenAI is definitely having a bad time.
This. Saturated benchmarks alone are not a measure of true model capabilities. Anyone who would have used this in production by now should understand these.
The standard Artificial Analysis pallete of tests might be unfair to Neon, and there are aspects where it is *far* ahead of everything else on the field. This is not an exaggeration.
(some of this fanboying might come across as "promotional" but I am just happy using a great model and multiple Google people in my DMs are taking my feedback; I am still not 100% happy with agy cli and they are listening to all my feedback on it, on a weekend, this is a big deal!)
1
1
98
The changes are visible from last release. Very well done. the differences to opus 5.5 are very miniscule on in terminal bench and frontier code!
Lots of work we've done to make Antigravity the best harness for complex tasks using Gemini. Neon using AGY CLI is now #1 on AA Coding agent Index and a massive jump over mini-swe on SWE-sweep, a bug finding and fixing eval on large codebases. It's a much larger jump than the native harnesses for the other frontier models.
More improvements to come to the harness in the next week 👀
49
The first time Google model Neon is at par with Claude in terms of steerability and doing stuff well in agy harness!
If google can keep using TPUs and keep costs lower they will be more powerful than other frontier labs and Nvidia for sure!
Google suite + Gemini product integration is what they can start focusing on! They can also start lowering GCP vertex costs to hit at AWS pricings and cut some sizeable market share there.
5
8
342
huduga retweeted
I seriously think US open-weight models wouldn't be this good without Chinese open-weight models and vice versa.
And proprietary models wouldn't be this good (and cheap) without open-weight models.
Competition is a good thing.
huduga retweeted
Reset everyone’s Gemini quota on Antigravity.
The caching improvements have increased the amount you can get done in quota by 22%. Enjoy building with Neon over the weekend!
Gemini 4 Neon just made Thinking Machines' Inkling 2 Mini run almost 6x faster on AMD hardware, by writing and optimizing custom GPU kernels on its own.
An AMD engineer gave Gemini 4 Neon the Inkling 2 Mini GGUF and let it handle the entire process on a Strix Halo (8060S) GPU: writing kernels, profiling them, finding bottlenecks, optimizing and benchmarking.
It worked for 4.3 hours and made around 490 tool calls.
The final results compared to llama.cpp on ROCm:
Decode: 44.8 tok/s vs 23.8 — 1.88x faster
Prefill: 1,465 tok/s vs 233 — 5.9x faster
4-stream aggregate: 100.8 tok/s vs 54.1 — 1.86x faster
The generated kernels reached around 80% of the GPU's theoretical memory bandwidth.
An AI model autonomously writing and optimizing low-level GPU code to get this kind of performance is genuinely impressive.
So Sarvam no longer has subsidy from government after March 2026! for GPUs ?
Replying to @cneuralnetwork
Government has spent tax money on private company, they can't escape without answering
moneycontrol.com/artificial-…
1
1
178
huduga retweeted
Summary of a medium-sized job (update a library, import that, re-implement some features, release new version of package etc) from Neon.
Opus, Fable, Astra all of them would not write with this level of information density and instead vomit multiple verbose paragraphs.
I do really like this ! So easy (and fast) to read it.
When you are coding and working in abudant AI era, there is lots of tasks and features that zip past your brain. But you have to delegate it until it is verified one by one.
Also, the outcomes delivery expectations timings are shrinking; this is bound to cause insane fatigues over the time if not managed well.
#observations
2
134
huduga retweeted
Today we're launching @arcfront with @Naman_Bhalla and deep support from @asxna.
In under a year, four of the five frontier labs became customers.
In 2025 an AI solved all 12 problems at the ICPC World Finals. The best human team solved 11. I competed in two of those finals, so I know how big a feat that is.
And yet, put that same AI inside a company as a finance manager, or in almost any other job, and it fails on day 1.
Why is that?
This is the stark reality; the issue is lot of engineering colleges are milking money and not using their money to improve and upgrade research acumen in the students. They need to look at organisations @lossfunk and china's universities on how to build them instead of doing nonsense like Galgotias!
Yesterday, I met a student from a reputed engineering college at our booth in BTS. He told me that, for the first time, many students in his college are yet to get placed. Campus placements have dropped sharply this year.
He was not sure if this was due to the impact of AI. But he seemed worried.
He regularly watches our podcast and asked formyy take. Should he continue looking for a job or consider entrepreneurship? He has a few ideas of his own.
I shared a few thoughts with him.
I strongly believe in promoting entrepreneurship. But I also believe we should be careful about encouraging everyone to take that path. Someone who has difficult financial circumstances or family responsibilities may not be in a position to take on the uncertainty that comes with entrepreneurship. It is not right to push someone into that situation just because we want to promote entrepreneurship. As we say in Kannada, ಕಂಡೋರ್ ಮಕ್ಕಳನ್ನ ಬಾವಿಗೆ ತಳ್ಳೋದು ಸರಿಯಲ್ಲ!
1. First, look at your financial responsibilities.
Entrepreneurship looks glamorous from the outside. But it is a demanding mental game.
I asked him if his family depended on him financially. Did he have an education loan or other financial obligations?
If yes, my advice would be to find a job first. Fulfil these responsibilities and build some financial stability. Then consider entrepreneurship.
2. Learn to operate in uncertainty.
Entrepreneurship means operating in a world of unknowns. There is no clear roadmap. You often will not know what to do next. Plans change. Things go wrong. You have to make decisions with incomplete information.
If you always need clarity, certainty and a clear plan, entrepreneurship can be difficult.
This is something you can learn. But you must be willing to operate in uncertainty.
3. Be prepared to unlearn.
Entrepreneurship will take you out of your comfort zone.
You may have to move away from familiar surroundings, spend less time with friends and make choices that people close to you may not understand.
You must be willing to question what you have learned and develop new ways of thinking and working.
4. Develop a high tolerance for pain.
You can put in your best efforts and make what you believe are the right assumptions. But things may still not go as planned.
You cannot blame your customers, team or external factors every time something goes wrong.
You must take responsibility, accept setbacks and keep going.
If the thought of dealing with repeated setbacks frightens you, perhaps you are not ready for entrepreneurship.
5. Be willing to learn and adapt.
You do not need all these qualities on day one. Nobody does.
What matters is whether you have the mindset to learn, adapt and improve. Do not spend too much time analysing every possible outcome. At some point, you must take action.
6. Do your homework before you start.
Understand the problem you want to solve.
Use AI tools to research it. Find out how many people face this problem and whether it is severe enough for them to pay for a solution.
Run surveys. Talk to potential customers. Build a small product and test it.
Do not spend months building something based only on assumptions. Look for evidence that people are willing to pay.
Once you have validated the problem and have some early traction, there is plenty of support available through grants, government programmes, early stage investors, angel networks and mentors.
My overall advice was this:
Be a self starter. Have a high degree of personal initiative. Be comfortable with uncertainty and develop a high tolerance for pain.
If you can develop these qualities, entrepreneurship is definitely worth trying.
He thanked me and seemed happy with the chat.
I also told him a thing at the end.
What I told him might give him a temporary high.
Give it 2 days. Think about everything we discussed. If you still believe these are challenges you are willing to take on, then give entrepreneurship a serious try.
huduga retweeted
Adani CFO understands AI better than most AI bros.
Firstly, from @AdaniOnline congrats to @AnthropicAI
Second,
Cross-Regime Category Error: Why Legacy Bankers Can’t Price Synthetic GDP
Comparing Anthropic’s $1.5T valuation to India’s top 10 corporates as "undervaluation"vs. "catching up" is a textbook category error in asset pricing. It benchmarks a convex call option on synthetic Labor in a new economic substrate against discounted cash flow (DCF) claims on a human-constrained economy.
1. The Ceiling of Human-Economy Equities
Traditional blue chips (banking, energy, staples) are linear yield vehicles anchored to biological human GDP. Growth is capped by demographics, credit penetration, physical consumption, and waking hours—priced via mean-reverting multiples (P/E, EV/EBITDA).
Calling them "undervalued" next to frontier AI is like calling Mumbai real-estate undervalued next to the discovery of thermodynamics.
2. Anthropic as a Call Option on Synthetic GDP (Yₛ)
Cognitive labor is decoupling from human population growth. Frontier models aren't SaaS tools; they are the "infrastructure of intelligence" for a parallel synthetic workforce (Yₛ) operating above human output (Y_Total = Yₕ + Yₛ). At $1.5T, markets aren't capitalizing EBITDA. Using Black-Scholes intiution Wall Street is pricing implied volatility and right-tail convexity of a 24/7 digital workforce scaling at the marginal cost of compute.
@AdaniOnline , we are already operationalizing this reality through our cobotic workforce model—a dual-track organizational architecture weaving human creativity with autonomous AI agents and robotics across our utilities, ports, and operations, upskilling 11000+ professionals annually to lead a ybrid industrial paradigm.
3. The 1999–2020 Precedent: Euphoria Still Underprices Regime Shifts
In 1999, Mary Meeker and Dow 36,000 were mocked as peak-bubble delusion when global e-commerce was ~$150B, internet penetration was 5% (277M users), and $1T tech firms seemed impossible. When the NASDAQ peaked at 5,048 in March 2000 and crashed 78% (wiping out ~$5T and bankrupting Pets.com and Webvan), legacy finance declared the thesis dead.
Yet by 2020, reality crushed 1999’s wildest models: 4.5B+ internet users (57% of humanity), $4.2T+ in annual e-commerce, >60% digital ad share, and core winners (Apple, Microsoft, Amazon, Alphabet) compounding into a $16T+ asset class. Legacy allocators mistook the 2000 clearing event for a ceiling, missing a new digital GDP layer.
4. The Epistemological Blind Spot of Legacy Bankers & Capital
This is classic lender-mindset orthodoxy: mistaking backward-looking accounting ratios for capital creation. Allocators trained in capital-scarce banking worship trailing multiples, static book value, and incremental yield—dismissing any paradigm-building capex cycle that breaks 20th-century spreadsheet as a "bubble."
While legacy finance debates scoreboards in the rear-view mirror, we have laid out our $100B+ (₹10 lakh crore) roadmap for the "Infrastructure of Intelligence"—mastering the full physical-to-digital stack from Port to Power to Metals to Processor, integrating gigawatt-scale renewable energy, grid transmission, and hyperscale liquid-cooled AI data centers (scaling toward 5 GW). Because intelligence is written in code, but it runs on electrons and sovereign infrastructure.
India’s blue chips are neither undervalued nor "behind"—they are yield vehicles monetizing the physical human economy, while those building the energy-and-compute backbone are underwriting the new substrate of an add-on synthetic economy. You cannot measure a non-linear dimension with a static ruler built for the old world ( much like 2000, clearing out will occur here too). Scientifically, Ai is a bigger shift, akin to development of scientific method in 16th century thus an order of magnitude a greater opportunity.























