Sabitlenmiş Tweet
Ujjwal
6K posts
Ujjwal
@dillisingh
I run a B2B SaaS startup @harmonizehq and tweet about startup life and practical wisdom in books.
Delhi Katılım Aralık 2013
535 Takip Edilen2.2K Takipçiler
its mind boggling
..but so much of this stems from standard accounting applied in the token world. as an AI firm, if I charge 20$ to the user and I in turn pay 8$ to anthropic who in turn pays $3 to nvidia/goog for cloud/hardware, total revenue claimed is 20+8+3
that said even with a 40% haircut.. 18B of real value created in such a short time which makes it historic
English
Anthropic is on a tear. $6BN --> $30BN in less than a year... unparalleled revenue ramp #GOAT
Which brings up the other imp question👇
How much of this $30BN is being double counted by its downstream token/API consumers?
Claude tokens are consumed on its own platform, by enterprises / SAAS cos and videcoding tools - Cursor, Lovable, Replit, Bolt etc. The vibe coding tools essentially are resellers (aka API wrappers) of tokens with some value added on top, but their claimed ARRs includes token cost paid to Anthropic.. apart from subscription revenue.
If Cursor is on a $2BN run rate, and Lovable $400MN, clearly a big part of their claimed ARR is just tokens, and hence being double counted here?
How should we properly read vibe-coding ARR numbers? Feels like the same AI revenue pool may be getting counted multiple times across app, infra, and model layers.
What’s the cleanest framework to think about total AI revenue pool?
Anthropic@AnthropicAI
Our run-rate revenue has surpassed $30 billion, up from $9 billion at the end of 2025, as demand for Claude continues to accelerate. This partnership gives us the compute to keep pace. Read more: anthropic.com/news/google-br…
English
Two key areas where using coding agents have proven to be a blessing to our team
1. automating tasks that just never got automated because we were always going to "write a script to do this the next time I have to do it.."
2. some things got tested on staging/prod where setting up a test harness on local was too darn hairy or cumbersome
Never thought shell scripts and TUIs will rule 2026
English
Unrelated but I still remember my time spent with English colleagues (during my team at Cricinfo).. we supported the same team during a Premiere League game and then stood in Chinnaswamy wearing opposing jerseys a few hours later..it was confusing and mind expanding at the same time..
Ujjwal@dillisingh
almost every opinion (negative) i've had of Aussie cricketers, growing up as a 90s Indian cricket fan, has been challenged in the past decade. growing up, is mostly about realising so much of your "beliefs" come from clashes of culture borne out of distance.. you shatter that distance & spend a few seconds with each other and boom.. it all dissipates quietly!
English
almost every opinion (negative) i've had of Aussie cricketers, growing up as a 90s Indian cricket fan, has been challenged in the past decade.
growing up, is mostly about realising so much of your "beliefs" come from clashes of culture borne out of distance..
you shatter that distance & spend a few seconds with each other and boom.. it all dissipates quietly!
Gagan Chawla@toecrushrzzz
From showing Javagal Srinath the bat and asking him to fk off !!!! From shoving off Sharad Pawar from podium after Champions Trophy win… From the controversial Sydney Test of 2007.. To his family treating Khaleel Ahmed as their own… To consoling likes of Prabhsimran and Nehal as a father figure after losing the finals last year!!! To saying ‘ ਆ ਗਇਆਂ ਹਾਂ ਭਰਾਵਾ’ IPL has changed the landscape of cricket enormously and Ponting has literally become a guiding force for every Punjab Fan!!! ਜੀ ਆਇਆਂ ਨੂੰ ਕੋਚ ਪੰਟਰ ਸਾਬ!!
English
Every side sounds like it’s winning 😅
Reminds me of World War II in Color documentary on Netflix where multiple countries were confidently reporting victory… at the same time.
Maybe just read the boring newspapers that wait. Less entertaining but far less anxiety inducing.
English
Being able to run small evals and trainings on datasets on our local mac machines before setting it all up on the cloud is proving to be a big productivity boost to us 👏👏
We've made functionforge in case you want to eval->finetune->eval a model for function calling. It now support mlx-tune
github.com/AnaekBackend/f…
Abdur Rahim@_ARahim_
Yaay! 🎉 4k+ downloads and 460+ stars! Building this has been a wild ride. If you have an Apple Silicon Mac and want to fine-tune LLMs locally without changing your original Unsloth code, come join the party.
English
@volklub @mufaddal_vohra yeah Mitch's reaction was over the top not doubt but Arsh still should have raised his hand or offered some apology. he just walked back which made it appear that the throw on M was on purpose.
English
@dillisingh @mufaddal_vohra Mitch didn’t give him time to apologise, he got agitated.
English
Arshdeep Singh fined 15% of his match fees for inappropriately throwing the ball at Daryl Mitchell.
English
@volklub @mufaddal_vohra fair i think.. while the throw was on, Mitch was totally away from the stumps and Arsh threw it directly at him. Even if we concede it was a mistake, Arsh should have apologised immediately. He got carried away with Mitch's reaction.
English
here's the demo x.com/dillisingh/sta…
Ujjwal@dillisingh
Experimenting with Pi coding agent for structured brainstorming and built an extension that lets me: • branch ideas • label interesting threads • run /obsidian-takeaway to distill the branch • automatically generate an Obsidian MoC + idea notes so pi for exploration/thinking and obsidian for safe keeping. One small UX thing I noticed: /tree navigation can feel a bit flaky when selecting branches. Curious what you think @badlogic
English
.@kepano Obsidian CLI integration question here.
in AI-agent subprocesses, 'obsidian files folder="003 Ideas"' returned banner-only + exit 0.
Setting macOS TMPDIR + retries fixed it but the curious thing is the same command in terminal worked without a snag.
Is this expected IPC/runtime behavior? Any recommended machine-checkable success signal?
English
Experimenting with Pi coding agent for structured brainstorming and built an extension that lets me:
• branch ideas
• label interesting threads
• run /obsidian-takeaway to distill the branch
• automatically generate an Obsidian MoC + idea notes
so pi for exploration/thinking and obsidian for safe keeping.
One small UX thing I noticed: /tree navigation can feel a bit flaky when selecting branches.
Curious what you think @badlogic
English
We ended up finetuning functiongemma @UnslothAI with our actual production data. Small models improve significantly with the right dataset + a good eval harness
My guess is tool calling will quietly become the default for chat or consumer interfaces..
Ujjwal@dillisingh
Really exciting how much can you do with small models. I made a synthetic dataset & finetuned functiongemma using unsloth for a home automation use case and its amazing how good it got so quickly for those interested, the dataset & notebook are here kaggle.com/datasets/ujjwa…
English