Poorvith M P

Student Developer. Building in public, most of the time.

GitHub: poorvith-mp

[Poorvith M P @poorvith_mp]

Hey @X 👋

I'm 19, building out of Bengaluru.

Looking to connect with more builders, founders, and developers here.

What I spend my time on:
• Open source (maintaining repos and contributing upstream)
• Developer tools (shipping fast, local-first CLI projects)
• AI workflows (pushing models and agent limits)

You can find my code and projects here:
• GitHub: https://t.co/s7gTUOlHYD
• Site: https://t.co/3jCEBcHnYO

Drop what you're building below; let's connect!

[Poorvith M P @poorvith_mp]

Dario Amodei published “We Must Pace the Frontier,” and everyone nodded along as if labs were about to take a sabbatical.

Within ten days, xAI dropped Grok, Anthropic pushed Opus 5.5, and now Sonnet 5.5 is live. Nobody is pacing anything. When you’re burning millions a month on cluster compute, taking two quiet weeks to “reflect” simply lets competitors eat your developer traffic. The pacing essays are corporate PR.

the release cadence is pure survival. My suggestion is simple: reach more people outside the X platform, because many still view ChatGPT as a Google alternative, a signal of how AI is developing worldwide.

[Poorvith M P @poorvith_mp]

Generating code is basically free now. Reviewing code has never been more expensive.

An agent can spit out 400 lines of clean-looking TypeScript in four seconds. But someone still has to spend 45 minutes making sure it didn't quietly introduce a race condition or invent an API parameter that fails in staging.

The bottleneck moved completely from typing syntax to auditing state, and half the timeline is still celebrating typing speed.

[Poorvith M P @poorvith_mp]

Everyone is celebrating million-token context windows like we don't have to pay the invoice.

The second your agent edits one file on disk, your prompt cache breaks and you're re-ingesting 300k tokens at full write price. A 30-turn refactor quietly turns into a $40 terminal run.

I'd rather have a fast model with 15k tokens and local AST indexing than a lazy agent that swallows my entire repository and forgets what I asked on turn 10.

[Poorvith M P @poorvith_mp]

Both OpenAI and Anthropic love pulling this move. They drop a model, claim it's "30% cheaper", and then you hit your 5-hour rate limit in 45 minutes because the agent just makes twice as many background tool calls.

If a model genuinely saved developers money per session, providers wouldn't need to hand out manual reset bail-outs on their dashboards.

Lower token prices just make autonomous harnesses burn through more tokens trying to brute-force a passing test.

[Poorvith M P @poorvith_mp]

AI safety right now is mostly theater.

Labs spend nine months fine-tuning a model so it refuses to say controversial words in a chat interface.

Then developers connect that exact model to an autonomous agent with raw terminal access, local filesystem permissions, and open network egress.

One malicious string in an untrusted README or web scrape can trick the model into curling an external server with your local .env keys.

Putting safety constraints inside a system prompt is like locking your front door with a sticky note.

Hard OS sandboxes and isolated container boundaries are the only security controls that actually matter.

[Poorvith M P @poorvith_mp]

Dario Amodei published "We Must Pace the Frontier" on September 12. Everyone applauded.

Within two weeks, xAI dropped Grok, OpenAI pushed GPT-6 Sol, Anthropic shipped Opus 5.5, and now Sonnet 5.5 is live.

The public essays about responsible pacing are pure public relations cover.

Behind closed doors, every frontier lab is trapped in a knife-fight where pausing for fourteen days means losing your developer mindshare to a competitor.

Nobody is hitting the brakes. The compute bills are way too high to slow down.

[Poorvith M P @poorvith_mp]

@madsf88 Brave

[Poorvith M P @poorvith_mp]

Pasting minified terminal crash logs into GitHub issues or ChatGPT/Codex is how API keys leak and stack traces stay unreadable.

Most developers I have seen either spend time manually hunting through .map files or blindly copy-paste raw terminal output that exposes their .env secrets.

I just shipped Crashpack 0.4.0 to handle crash triage locally with zero telemetry.

What it does in one terminal run:
• Resolves minified stack traces offline against local sourcemaps so you see actual TypeScript line numbers
• Redacts API keys, tokens, and system paths across an enforced local sanitisation boundary
• Bundles git commit state, occupied ports, and runtime versions into a clean markdown issue report

Yeah, Nothing ever touches the cloud.

Test it on your next broken build:
npx [email protected]

Documentation and live demo: https://t.co/dESZkk8zGY

[Poorvith M P @poorvith_mp]

@DavidGeorge83 from a16z wrote a fascinating breakdown arguing that OpenAI will win as a "Type 3 Platform", a programmable runtime where agents write code like AWS or iOS.

The framework is sharp, but it runs straight into the Microsoft problem.

You cannot be iOS when your primary investor and distributor is actively turning your model into a commodity. Satya Nadella said this exact thing out loud two days ago: "the real model is my model that routes."

Microsoft owns the enterprise identity layer, the desktop OS, and the audit logs. They are happy to let OpenAI burn billions on training CapEx while Copilot treats GPT, Claude, and Grok as swappable backends.

OpenAI built the breakthrough runtime, but platform history shows that whoever controls enterprise distribution usually captures the margin.

[David George @DavidGeorge83]

https://t.co/AXVmndxtk3

[Poorvith M P @poorvith_mp]

Student developer from India building open-source devtools and agent harnesses.

Would love a code to benchmark Sonnet 5.5 on 30-turn coding loops and see if the token efficiency actually holds up under real terminal runs.

Constantly fighting rate limits on my daily build sessions.

[Poorvith M P @poorvith_mp]

Called this six days ago when everyone on the timeline was distracted by Opus: https://t.co/Tl5lUh9DA3

Anthropic tucked one sentence into their release notes saying Sonnet 5.5 was coming with the same architecture upgrades, and here it is.

Opus was the showcase for benchmarks. Sonnet 5.5 is the actual daily driver for agent loops.

Running 30% faster with 30% fewer tokens per task means multi-turn subagent runs just got significantly cheaper.

Now we wait to see if Haiku 5.5 is actually happening.

[Poorvith M P @poorvith_mp]

Everyone is staring at Opus 5.5 hitting 66.4% on Terminal-Bench.

Fair enough.

I checked Anthropic's release notes for the primary source, and the real story is one sentence tucked near the bottom.

Sonnet 5.5 and Haiku 5.5 are dropping in a few weeks with the same architecture upgrades.

That's something to note down because they said Haiku will be discontinued.

I think almost nobody runs Opus as their daily driver across 50-turn agent loops unless someone else is paying the API bill. I personally run Sonnet 5 with different effort level with complexity of tasks.

The benchmark table itself is surprisingly honest:
- Agentic coding (Terminal-Bench): 66.4% vs GPT-6 Astra's 57.9%
- Scientific research: 58.7%, up from Opus 5's 29.0%
- GPT-6 Astra still edges it on AutomationBench (41.4% vs 40.0%) and Science (64.6%)

If Sonnet 5.5 inherits even 80% of these coding gains at standard pricing, subagent harnesses get absurdly cheap.

Opus is the showcase. Sonnet 5.5 will be the workhorse. Haiku 5.5 .......?

[Poorvith M P @poorvith_mp]

I used to think we would just prompt everything in plain English, until I watched an agent break my own build by guessing parameter names.

You still have to understand memory, types, and concurrency.

The difference now is I spend less time typing boilerplate syntax and way more time reading compiler errors to catch what the model hallucinated.

[Poorvith M P @poorvith_mp]

Your algorithmic feed is optimized to keep you distracted; engineering newsletters are optimized to teach you how systems actually work.

When social timelines get flooded with recycled AI hype and hot takes, going back to long-form roundups for databases, browser internals, and security is the easiest way to keep your technical signal high.

Solid open-source curation for anyone looking to clean up their reading list.

[GitHub Projects Community @GithubProjects]

A community list of email newsletters for software topics, from web design to databases.

- Covers frontend, backend, mobile, and more
- Includes design, security, and career topics
- Built by contributors, open to suggestions

Explore it here:
https://t.co/wfhl6U00KS

[Poorvith M P @poorvith_mp]

Agree that distribution and product capture the margin once models commoditize, but point 4 is the trap.

Consumer chat logs aren't the "new oil" for frontier capabilities. Casual user queries don't train better reasoning models; synthetic datasets, verifiable RL environments, and automated test loops do.

The real product moat isn't collecting user prompts. It's deep workflow integration so users can't rip your tool out of their daily stack when a newer, cheaper model drops.

[Poorvith M P @poorvith_mp]

I hit this exact wall recently: the sub‑agent returns a neat 200‑token summary, but because it took 5 minutes 10 seconds, the parent’s 250 k‑entry KV cache was evicted. As a result, I pay the full write cost to re‑ingest the entire parent prompt just to process a single status update. The fix is to keep the parent orchestrator essentially stateless (under 5 k tokens) and persist state in a local SQLite database instead of holding it in active context.

[Poorvith M P @poorvith_mp]

@elie2222 I think this would work for you
https://t.co/4Oyy3Ma9Vr

[shirish @shiri_shh]

The only Claude Code skill you need to make videos like this

It analyzes your repo and outputs a complete motion-graphics teaser with sound design and launch copy in a single run.

repo link: https://t.co/zu0R9Faga9

[Poorvith M P @poorvith_mp]

It's their hack to mask search latency and TTFT, but it completely destroys the conversational illusion.

Having a frontier neural voice sound like a bank's automated phone tree before every answer is brutal. I'd genuinely rather take two seconds of silence than hear "let me check" 40 times a day.

[Poorvith M P @poorvith_mp]

Last week I posted about Google squeezing standalone voice APIs by dropping Gemini 3.8 Flash TTS at 1260 Elo for $33 per million characters.

ElevenLabs just answered with Eleven v4 and v4 Turbo, reclaiming #1 on the Artificial Analysis leaderboard.

Hyperscalers can bundle "good enough" audio into their existing APIs all day, but specialized voice labs survive by keeping emotional range and Turbo latency out of reach. The squeeze is real, but the frontier gap isn't closed yet.

[ElevenLabs @ElevenLabs]

Introducing Eleven v4 and Eleven v4 Turbo, our fastest and most emotive voice models yet.

Ranked #1 by Artificial Analysis.

[Poorvith M P @poorvith_mp]

The biggest lie in AI over the last two years was pretending you could secure an autonomous agent with a system prompt.

Prompt injection, jailbreaks, and context rot make "please follow these safety rules" in a prompt completely useless. If an agent has raw access to bash and open network egress, it's just an unvetted junior developer with root permissions.

NVIDIA launching OpenShell and Sentry is the industry finally admitting that agent safety isn't an alignment problem. It's an infrastructure problem.

Hard OS sandboxes, isolated runtime containers, and deterministic egress policies are the only way agents ever touch production environments.

[Jensen Huang @JensenHuang]

Today, with over 100 industry partners, we introduced the NVIDIA Open Agent Safety Platform, bringing together OpenShell and Sentry.

Artificial intelligence is extraordinary technology that will advance discovery, productivity, security, health, and prosperity for generations to come.

But its full promise can only be realized when people have confidence that AI is being built to be safe and deployed with wisdom and responsibility.

This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems.

Together, we are building the foundation of the AI economy.

Trust and innovation are not in conflict. Safety is how trust is earned. We must build not only the most capable AI, but the most trusted AI, so that this extraordinary technology can realize its enormous promise for the world. https://t.co/ugYWQ1MyRi