Ankesh Bharti
@_feynon
Followers
390
Following
30K
Media
530
Statuses
6K
26, applied researcher & technologist. on device model maxxing. building https://t.co/25TLHHPpct. publishing https://t.co/PJ38iC4XaW. community https://t.co/7tTyptKfRX.
Bengaluru, India
Joined May 2018
the constrained environment of creative tools created by non creatives because creatives refuse to create creative tools because its not creative enough. (Its changing tho.)
0
0
2
While everyone was distracted by flashier launches, @AnthropicAI quietly shipped a powerful memory tool in their SDK. This makes them the first model provider to take an opinionated stance on how memory should work. And it's really good! Blog, demo, and repo link ⬇️
26
34
536
ChatGPT Atlas builds on Chromium in a unique way to deliver on our ambitious product goals - enabling a fast, responsive browser with agent DNA. Read more about how we did this in a new engineering blog post by Ken Rockot & I.
openai.com
Inside our new process architecture, which gives you a faster, smarter way to use the web.
19
25
274
Identity and memory are two sides of the same coin. Tiles makes that coin yours: your user-agent.
0
1
1
Prediction: Meeting bots will go extinct. Local AI Notetakers will take over. @meetgranola started this wave but they are not taking advantage of being a local app yet, e.g. still uploading raw audio to a third-party cloud. As @tobi said, @tryhyprnote (YC S25) is great because
I’m actually very pro meeting recording and ai summarization. But I’m not ok with bots joining as fake humans accomplish this. It’s a meeting between you and me. Not you and me and some startup’s viral growth strategy. Granola is great. Gemini does this well in Google Meet.
3
1
8
The best AI roadmap is shaped by real builders. For weeks, one request kept showing up in our inbox: “Can Nexa support Qwen3-8B on Qualcomm NPU?" So we did. Qwen3-8B now runs fully on-device on @Qualcomm Hexagon NPU through NexaML. And today, we’re making this loop official:
3
13
63
More for less: the average paying Amp user now pays ~21% less per day vs. a few weeks ago, while # {threads,messages} {,per user}, users, & total revenue are all growing. We did this while keeping Amp's `smart` mode quality the same or better, and without subsidizing you (which
8
7
68
We now support Bun on Vercel. Bun is an open runtime pushing the frontier on DX and performance. Start building and see our latest benchmarks. https://t.co/WvE5sQzxe0
vercel.fyi
Vercel Functions now supports the Bun runtime, giving developers faster performance options and greater flexibility for optimizing JavaScript workloads.
56
97
1K
Tiles exists to make this possible: memory that stays private on-device yet connects across agents through open protocols.
The Memory Walled Garden. Our latest post explores why portable memory for AI would be good not only for users but also for application developers, and even the major model labs. Consistent with our idea that open protocols create an “architecture of participation” that enable
0
1
1
New physical attacks are quickly diluting secure enclave defenses from Nvidia, AMD, and Intel
arstechnica.com
On-chip TEEs withstand rooted OSes but fall instantly to cheap physical attacks.
0
2
3
After seeing @karpathy nanochat, I became obsessed with the idea of using GRPO fine-tuning on it locally on Mac. I got it working and open-sourcing the whole thing here Here's nanochat going through the steps of solving GSM8K dataset, altho may be too much for it 😅
8
12
150
Apple GPUs are much more powerful with the new Neural Accelerators in A19/M5+. But this upgrade did not significantly improve performance per watt (think sustainability for long-running tasks) to be competitive with the Apple Neural Engine. Given that 2025 was a dormant year for
@ggerganov My post is unclear. Neural Accelerators get throttled pretty heavily under load for my iPad Pro. It just fares much better than it is on A19 Pro, and for performance work I am doing, that makes a difference (on turnaround for testing). I don't think thermal is any better with NA.
5
4
73
We use Swift virtually everywhere at @argmax, including as the core of our Node and Python packages. Given the recent progress on Swift for Android, we will test how feasible it is to build the core of our Android SDK in Swift and report back.
📣Announcing the first preview releases of Swift for Android, enabling you to build Android business logic with the same Swift that you use for Apple platforms. https://t.co/UAR6LO3prQ
#Android
8
5
191
benchmarking inference servers is, as the kids say, *non-trivial*.
3
1
34
There are a sophisticated set of enabling technologies, but the feature’s goal was incredibly simple: to capture the vitality of the world around us. The Live Photo patent has a picture of my daughter as a toddler spinning in a dress in front of our house. As with a Live Photo,
live photos hit this perfect intersection of tech & tenderness. it might be the most quietly brilliant thing apple ever made. it’s a subtle wordless trick that turns stillness into something breathing. i just stumbled on one & it hit like a gut punch. ridiculous that all photos
16
40
833
Now in research preview: gpt-oss-safeguard Two open-weight reasoning models built for safety classification. https://t.co/4rZLGhBO1w
openai.com
New open safety reasoning models (120b and 20b) that support custom safety policies.
162
304
2K
