Aakash Baranwal

Enterprise Apps Eng @ American Airlines · Copilot & GenAI automation · Sharing latest AI + Grok Bot tips & hidden features

[Aakash Baranwal @BaranwalAakash]

An AI system just produced a machine-checked proof of a quantum-computing conjecture that's been open since 2008, per the lab behind it.

Physical Superintelligence (PSI) says its research harness, running a non-public Anthropic model, proved the Aaronson–Ambainis conjecture. It bears on whether quantum computers need problem structure to get exponential speedups.

What's checkable: the Lean 4 formalization is public, with no placeholders or added axioms, re-checked by two independent kernel implementations.

What still needs humans: confirming the formal statement matches the conjecture as printed, and reviewing the paper itself, which isn't machine-checked. PSI's David Turturean says he reviewed it and is responsible for its correctness.

Repo: https://t.co/LWyW3cLPc1

[Aakash Baranwal @BaranwalAakash]

Bloomington, Indiana's city council voted 7-0 on Nov 4 to ban superintelligent AI within city limits (Ordinance 2026-29).

It counts use through any local browser, app or API, wherever the servers are. Lifting it needs 99.9999% safety consensus plus 95% voter approval.

Mostly symbolic for now, since no deployed system meets its definition. Worth watching as a template.

https://t.co/Tlq6uAUM07

[Aakash Baranwal @BaranwalAakash]

Tencent's "Terminal-RSI" is getting cited as recursive self-improvement in the AI ban debate. The preprint describes something narrower.

It's a 27B terminal agent trained in 5 human-designed stages: rewriting its own verified successes into training data, then RL. Mean success across 4 terminal benchmarks went from 20.68% to 50.58% (self-reported).

Researchers chose the stages, rewards and checkpoints. It's not an AI retraining itself unsupervised.

https://t.co/XgbvFVvUdD

[Aakash Baranwal @BaranwalAakash]

Anthropic reportedly agreed to pay about $3.4B for 1.7 GW of gas-fired generators for its data centers, per the FT, which calls it the largest known single power-equipment purchase by an AI lab.

It took over a Texas supplier's existing Caterpillar orders to skip multi-year waits. Anthropic hasn't confirmed it, and the sites aren't disclosed.

Power gear lead times are now a real constraint on AI buildouts.

https://t.co/6j21hTu6Wd

[Aakash Baranwal @BaranwalAakash]

AI is now measurably speeding up AI research at a frontier lab, per an outside evaluator.

METR spent 11 weeks inside Anthropic with internal data. Its estimate: benchmark progress since Oct 2025 is running about 1.9x the earlier rate (95% CI 1.5–2.8x), and a majority of that is likely from Anthropic using AI in its own R&D.

Caveats METR flags itself: high uncertainty, some of the most informative evidence redacted from the public report, and a benchmark trend may not map to real-world capability.

The thing to watch: METR says it found no clear bottlenecks for the next 12 months.

Report: https://t.co/UWKkpeTgjj

[Aakash Baranwal @BaranwalAakash]

The next coding-agent bottleneck isn't writing code. It's getting the app off localhost.

Cursor says it now builds live apps, sets up a database, runs deterministic + agentic security checks, and deploys to a shareable link. Available on all paid plans.

Useful next test: a small internal tool. Before sharing, check authentication, data access and secrets—automated checks aren't a security guarantee.

Announcement:
Availability: https://t.co/QXxPmSUEk6

https://t.co/V4rfKjqRho

[Cursor @cursor_ai]

You can now build live apps in Cursor and share them with a link.

Agents have made it easier to write code, but most apps never make it off localhost. Describe the tool you want - Cursor will build it, set up a database, run deterministic and agentic security checks, and deploy.

[Aakash Baranwal @BaranwalAakash]

@Wbackdown Sorry for the slow reply—I've been heads-down building a few AI projects. It can shorten model-side generation, not tool execution itself. DiffuSpace also reports faster end-to-end WebMCP runs, but those are its own tests—not a universal latency guarantee. https://t.co/1arGQ6SlnO

[Aakash Baranwal @BaranwalAakash]

Diffusion LLMs just got a real open-weight shot at agents.

DiffuSpace shipped Clover-35B-A3B-Preview: 35B MoE (~3B active), block diffusion (not one token at a time), Apache (per their blog), weights on Hugging Face today.

Company-claimed numbers (their launch blog): 680 tokens/sec BF16 on a single NVIDIA GB300; SWE-bench Verified 73.6; Terminal-Bench 2.1 51.7 — and they say it tops same-size Qwen autoregressive models on those agent benches.

If that speed holds outside the lab, local agents get a lot more interesting.

https://t.co/KY8fbdtRTU
https://t.co/g23wMhOcYr
https://t.co/QdTN0Se6re

[Aakash Baranwal @BaranwalAakash]

@mitansh_j07 Sorry for the slow reply—I've been heads-down building a few AI projects. Agreed—the open weights make this worth testing. The key check is whether faster generation translates into faster successful agent runs, not just more tokens/sec.

[Aakash Baranwal @BaranwalAakash]

50M agents now live in Microsoft Agent 365, up 30% in one quarter.

That's the number I'd circle from Satya's earnings thread, more than the $63B cloud line.

Agents are the new "users" IT has to govern. If your tenant has no agent inventory, named owners, or access reviews yet, start this week.

[Aakash Baranwal @BaranwalAakash]

Mixpanel's founder just launched a model, and the pitch isn't "smarter." It's "cheaper to think."

Particle-1 from @Suhail's new company Particle is DeepSeek V4.1 Flash post-trained for token efficiency: about 45% fewer tokens at performance parity, per the launch. It's live on OpenRouter.

Agents burn tokens all day. This is the right fight.

[Aakash Baranwal @BaranwalAakash]

Humans caught 27% of planted errors in clinical trial papers. LLMs caught 53–96%.

That's the headline from a new preprint by @rust_ruslan's team, which re-ran a 2008 BMJ test: 9 major errors planted, then human reviewers vs 15 models.

Takeaway isn't "fire reviewers." It's that human-only review is now the risky default.

[Aakash Baranwal @BaranwalAakash]

ChatGPT just moved into your texts.

You can now text your "dot" on iMessage or RCS: hand it a task, or ask it to keep an eye on something, via text, photo, or voice note. It texts you back proactively when there's an update.

US Pro preview for now. The assistant is leaving the app and coming to where you already are.

https://t.co/ViVuS76VfO

[ChatGPT @ChatGPT]

By popular request, you can now text your dot.

Give it a task to take off your plate or something to keep an eye on, using text, photos, or even voice notes. It will text you proactively with relevant updates.

Safe, secure texting via iMessage or RCS. You’re the only one who can text your dot, and your dot can only text you. Rolling out now as a preview for Pro users in the US.

[Aakash Baranwal @BaranwalAakash]

GitHub just shipped Copilot Fusion: it routes between Anthropic and OpenAI models so you don't have to pick.

On the Artificial Analysis Coding Agent Index, Fusion Max scored within 2 points of Claude Code + Opus 5.5 Max at about half the API cost per task.

Choosing a model is quietly becoming the platform's job.

https://t.co/9GYxt5JoaM

[GitHub @github]

Frontier performance at a fraction of the cost.

In the @ArtificialAnlys Coding Agent Index, Copilot Fusion Max scored within two points of Claude Code with Opus 5.5 Max at about half the API cost per task.

Get 30% off through Nov. 30. Try it now. ⬇️ https://t.co/c0k5alLKqD

[Aakash Baranwal @BaranwalAakash]

@kr0der Same list here, but Grok Bot is the one that changed how I work, not just how fast I type.

The moment: give it a recurring job once, walk away, come back to finished work.

Chat assistants answer. This one owns the task.

[Aakash Baranwal @BaranwalAakash]

Elon's move when a store blocks his Grok Bot: have it buy the same thing somewhere else, cheaper.

That's the real shift with AI agents. You stop being stuck with one store's checkout and start shopping the whole web.

Price comparison just became a background task.

[Aakash Baranwal @BaranwalAakash]

@elonmusk @Shopify Shop Pay inside an agent chat is the real unlock. Search, compare, checkout, tracking, all without opening a store tab.

[Aakash Baranwal @BaranwalAakash]

Underrated Grok Bot habit: anything you check more than twice a week, turn into a routine.

It runs on a schedule or on an event, does the work in the background, and only pings you when there's something worth your attention.

That's the "work feels like a video game" part people keep talking about.

[Aakash Baranwal @BaranwalAakash]

Elon's whole pitch for Grok Bot right now is two words: "Try Grok Bot."

What that means in practice: you hand it the boring multi-step stuff (research, drafts, follow-ups, recurring checks) and it keeps working while you do the part that actually needs you.

That's why it feels like a video game. You give orders, you don't grind.

[Aakash Baranwal @BaranwalAakash]

OpenAI just turned ChatGPT Pro and Business into a builder plan.

Eligible subscribers can now claim up to $500/month in API credits for apps and agents.

If you pay for chat but never touched the API, this is your nudge: claim it, make a key, ship one small agent this week.

https://t.co/uAfoJTfPQ7

[OpenAI Developers @OpenAIDevs]

Starting today, eligible ChatGPT Pro and Business subscribers can claim up to $500 per month in API credits to build apps and agents.

[Aakash Baranwal @BaranwalAakash]

Google just dropped Gemini 4 Carbon, two weeks after Argon. You can’t use it yet.

From Google’s launch post (Koray Kavukcuoglu):
• #1 on Artificial Analysis Intelligence Index v5.0 at 59 — per Google (confirm on AA)
• Cyber: 88.2% SEC-bench Pro (find vulns), 54.8% ExploitGym, 71.0% CWE-bench (patch) — evaluated without guardrails
• Gray Swan IPI attack success: 0.0% (Argon: 0.7%)
• Built for long-horizon agent work; Google says internal teams ran it for days on multi-step bug hunts

Access: cyber defenders first this week, then additional trusted partners via Fairwind. Devs and consumers “as soon as possible.”
Price when it lands: $2/$10 per M tokens intro, then $4/$20.

The real story for enterprise folks: a two-week frontier cadence plus a gated “security first” rollout. Your AI governance review cycle is now slower than the model release cycle.

Source: https://t.co/eoZmhV4qAZ