DEV Community: Abhishek Jambla

Grok Bot Can Now Make Phone Calls for You. Who Owns the Mistakes?

Abhishek Jambla — Mon, 09 Nov 2026 17:56:05 +0000

What’s new: xAI’s Grok Bot can now call a business, talk to a real person, and report back. Musk previewed it Oct 25. It rolled out in early November, for US numbers only.

How it works:

You say who to call and what you want, plus rules (“don’t accept a discount”).
A confirmation card shows who it will call; you press Call.
A live transcript runs, with Take over and End call buttons.
If it’s stuck, a question card asks you. You type, it speaks.
It must say it’s a virtual assistant, and calls may be recorded.

Why it matters:

One creator reportedly cut an internet bill from $84 to $45 a month. The bot paused for approval before accepting.
Not every call works: one rep refused to cancel once they realized it was an AI.
Per the terms (as summarized by AiNews), customers are responsible for the agent’s actions, even unintended ones.

Dev takeaway: Build agents with approval gates, clear permissions, and logs. Google’s 2018 Duplex also handed off to humans when stuck.

Tags: #ai #agents #webdev #productivity

Media

🖼️ Duplex flow diagram (user asks Assistant, Duplex calls the business): https://storage.googleapis.com/gweb-research2023-media/original_images/Duplex_assistant.png
🖼️ Duplex speech pipeline diagram (sound in, text analyzed, speech out): https://storage.googleapis.com/gweb-research2023-media/original_images/Duplex_ASR_system.png
🎥 Peter Yang’s real Comcast call with Grok Bot: https://youtu.be/F-356TczoQU

References & Credits

Teslarati
AiNews, Alicia Shapiro
Google Research, Yaniv Leviathan and Yossi Matias (images)

AI Found 948 Linux Kernel Bugs — Here's How OpenAI's Models Did It

Abhishek Jambla — Sun, 08 Nov 2026 09:34:39 +0000

Tags: #security #openai #linux #ai

OpenAI's o3 and o4-mini reasoning models just changed what we thought AI could do in security research. Working alongside researchers, they surfaced 948 CVE-grade vulnerabilities in the Linux kernel — and autonomously exploited one of them end-to-end.

The key finding — CVE-2025-37899:

A use-after-free (UAF) bug in the Linux kernel's ksmbd SMB server module
Exploitable remotely and without authentication
The models generated working proof-of-concept exploits that bypassed both ASLR and KASLR — standard memory-randomization defenses on modern Linux systems
How the AI pipeline worked:

Models analyzed existing kernel patches to find similar bugs in adjacent code (variant analysis)
No human wrote the exploit — the models did it autonomously
OpenAI responsibly disclosed to Microsoft; the bug is now patched
Why it matters for developers:

The time between disclosure and working exploit is shrinking fast
AI lowers the skill floor for exploitation — defenders need to patch faster
Same capability cuts both ways: OpenAI also used AI to find 20+ CVEs in open-source projects defensively
OpenAI launched a $1M Cybersecurity Grant Program in API credits for defensive research.

📎 Media:

🎬 OpenAI for Government Forum: Advancing Cybersecurity — Official YouTube
🔗 NVD CVE-2025-37899 detail page
References & Credits:

AI News — OpenAI cyber model found 948 Linux kernel exploits
OpenAI — Official cybersecurity page
TechCrunch — OpenAI models help find 948 vulnerabilities (July 2, 2025)
The Verge — OpenAI 2025 Cybersecurity Report
HackRead — AI Can Now Write Actual Exploits
NVD — CVE-2025-37899
Video credit: OpenAI (official YouTube channel)

Notion’s AI Agents: Your Workspace Now Does the Work

Abhishek Jambla — Fri, 06 Nov 2026 17:40:14 +0000

Tags: ai, productivity, automation, saas

Notion 3.0 (Sept 2025) turned its AI from a helper that suggests into an agent that acts, inside a workspace Notion says is used by 100M+ people.

How it works

Goal in, finished work out: it builds pages and databases, then updates them.
Long runs: up to 20 minutes of solo work across hundreds of pages.
Connected tools: it can pull from Slack, email and Google Drive.
Memory page: a plain Notion page of instructions you can edit anytime.
Your access, no more: agents follow your existing permissions.
MCP integrations: outside AI tools can plug in.

What came next

Custom Agents run in the background on a schedule or trigger. Beta started Feb 2026.
Users built 1M+ in two months.
Since May 4, 2026 they’re paid, via credits ($10 per 1,000) on Business and Enterprise plans.

Why it matters
Your docs and databases are already the context, so there’s no copy-pasting into a chatbot. Start with one boring, repeatable job like a weekly status report or ticket triage, and let an agent own it.

🎥 Video: https://www.youtube.com/watch?v=R1cF4T4lgI4

References & Credits

Notion, Introducing Notion 3.0, by Akshay Kothari (text, GIF and video © Notion)
Notion, What we learned during the Custom Agents beta
TechCrunch, Notion launches agents for data analysis and task automation, by Ivan Mehta
Datamation, Notion Unveils AI Agent for Workflow Automation
Trending Topics, Notion launches AI Agents, by Jakob Steinschaden (the 100M+ users figure)

FlashProve: Proving Which AI Model Actually Answered

Abhishek Jambla — Wed, 04 Nov 2026 06:37:42 +0000

You call an AI API. How do you know the model that replied is the one you asked for? Today, you don’t. Anthropic’s FlashProve (open source, Apache 2.0) is a research prototype that fixes this with a cryptographic proof.

How it works

Commit: The provider publishes a short “fingerprint” of its model weights. The weights stay secret.
Prove: For each reply, the provider’s GPU creates a proof that the committed model produced those exact tokens.
Verify: You check the proof on a normal CPU. No GPU, no weights needed.

Numbers (Llama 4 Maverick, 400B)

About 110 tokens/sec on one H100
Proof size ~19 MB; check time ~17 sec
Proving costs 70-95x the GPU time of normal serving

Why it matters

Catch secret model swaps by routers or providers
Spot tampered responses
Give auditors and regulators real proof

Limits

Only 1,024 tokens of context
Proves an integer version with greedy decoding
Hides weights, not architecture
Prototype: not maintained, no deployment plans

Media
Visual aid: Interactive Lean theorem visualizer from the FlashProve repo, which shows the security claims being checked

References & Credits

Anthropic, Progress on provable inference (Roy Rinberg, Alex Allain and co-authors from EPFL, MIT, StarkWare and Succinct)
FlashProve on GitHub (Apache 2.0)
TechCrunch, Russell Brandom
Interesting Engineering, Neetika Walter

Interseting

Abhishek Jambla — Sat, 31 Oct 2026 15:20:54 +0000

Abhishek Jambla

Oct 31

Gemini 4 Carbon: What Developers Should Know

#ai #gemini #google #agents

1 min read

Gemini 4 Carbon: What Developers Should Know

Abhishek Jambla — Sat, 31 Oct 2026 15:14:55 +0000

Google announced Gemini 4 Carbon on Oct 28, 2026, two weeks after Gemini 4 Argon. It ranks #1 on the Artificial Analysis Intelligence Index v5.0 (score 59), just ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57).

What’s new

Long-running coding: Google says Carbon worked for days on hard bugs in big codebases.
Security skills: 88.2% on SEC-bench Pro (finding flaws), 54.8% on ExploitGym (checking if a flaw can be abused), 71.0% on CWE-bench (fixing the root cause).
Web dev: +89 Elo, now #4 on WebDev Arena.
More honest: about 2x fewer unsupported claims than Argon.
Prompt injection: 0.0% attack success on Gray Swan’s benchmark.

Access and price

Today: select cyber defenders only, then more partners through the Fairwind Program.
Intro price: $2 / $10 per million input/output tokens, later $4 / $20.

Why it matters: An agent that stays on task and fixes bugs at the source could save real review time. Keep in mind most scores are Google-reported, and no technical report is out yet.

Tags: ai google cybersecurity programming

Media

🖼️ Security benchmark chart: https://storage.googleapis.com/gweb-uniblog-publish-prod/images/gemini_4_carbon_secbench-pro.width-1200.format-webp.webp (Carbon on SEC-bench Pro, vulnerability finding)
🖼️ Patching chart: https://storage.googleapis.com/gweb-uniblog-publish-prod/images/gemini_4_carbon_cwe-bench.width-1200.format-webp.webp (Carbon on CWE-bench, root-cause fixes)
▶️ Third-party explainer: https://www.youtube.com/watch?v=V-UtDrdOlR4

References & Credits

Google / Koray Kavukcuoglu, Gemini 4 Carbon: our next frontier model (charts: Google)
9to5Google, Engadget, VentureBeat
Reference article: AI Magazine

How Google DeepMind Uses Games to Build Smarter AI Agents

Abhishek Jambla — Sat, 31 Oct 2026 05:33:52 +0000

Google DeepMind’s Alexandre Moufarek (Director, Inception team) says games are still the best training ground for AI. The simple version:

What is SIMA 2?

An agent that looks at the screen and uses a normal keyboard and mouse, like a player
Needs no game code or special APIs
Powered by Gemini, so it can reason, explain its plan and chat

How it learns

Starts from human gameplay videos
Then practises alone, with Gemini giving feedback
Understands sketches, emojis and other languages

Why it matters

Task success rose from 31% to 65%
It handled worlds generated by Genie 3 that it had never seen
Game uses: smarter NPCs, AI companions, automated QA testing
Next: an offline EVE Online sandbox to study memory and long-term planning

Limits: still research. It struggles with long multi-step tasks and has short memory.

Media

🖼️ Chart comparing SIMA 1, SIMA 2 and humans on task success: https://lh3.googleusercontent.com/KwuGClFlsydWIS2azo2WjBWkMY-Hm3geOZAad00wP-CMtmWQa8DlpmqOMvXC6QKiPnEUBgtToVWnJ9rzz0O7dTcaVFhYOSggEugNVJBWVgc1hegDfg=w1440 (Image: Google DeepMind)
🎥 Official intro video: https://www.youtube.com/watch?v=Zphax4f6Rls

References & Credits

Google DeepMind – SIMA 2 (SIMA Team)
Google DeepMind – From Atari to EVE Online (Alexandre Moufarek & Adrian Bolton)
AI Magazine – reference article
Decrypt – SIMA 2 coverage (31% vs 65% figures)
MIT Technology Review – SIMA 2 and Gemini
Image and video: Google DeepMind

Amazon to Engineers: Stop Tokenmaxxing

Abhishek Jambla — Tue, 27 Oct 2026 19:05:43 +0000

What happened: Some Amazon staff chased high token counts (the chunks of text an AI reads and writes) because internal leaderboards rewarded usage over results. In May, Amazon shut one down. On Oct 27, a public memo went further.
CNBC

How it played out

Goal: 80%+ of developers using AI weekly.
tomshardware
Some staff reportedly used an in-house agent platform to run needless tasks and boost their numbers.
tomshardware
Every extra token is real compute cost.

What the memo says

Token usage, tool adoption, and code-change counts are “deeply flawed” as targets.
CNBC
aboutamazon
Judge AI by long-term outcomes: customers, business, team.
If AI doesn’t fit your task, say so and explain why. That’s judgment, not failure.
Used with care, usage metrics still correlate with output.

Why devs should care

Measure shipped value, not prompts sent.
AI writes fast; you still own quality, so review carefully.
Spec-first prompts (like Kiro’s requirements, design, tasks flow) cut ambiguity.

Tags: #ai #softwareengineering #productivity #career

Media

Image: Amazon News header graphic (official page image). https://assets.aboutamazon.com/dims4/default/7e16f44/2147483647/strip/true/crop/2000x1000+0+63/resize/1200x600!/quality/90/?url=https%3A%2F%2Fassets.aboutamazon.com%2F52%2Fa8%2F3c8ee3984dd9a2548f89d35d787c%2Famazon-logo-hero-images-2024-4.jpg
Video: Werner Vogels’ re:Invent 2025 closing keynote (AWS Events). It covers the “Renaissance Developer” idea and a Kiro spec-driven coding demo. https://www.youtube.com/watch?v=3Y1G9najGiI

References & Credits

CNBC: https://www.cnbc.com/2026/10/27/amazon-ai-coding-tokenmaxxing.html
Amazon News, “An update from Amazon engineering leadership…” (Terk Terkowitz): https://www.aboutamazon.com/news/company-news/amazon-software-engineer-role-guideline-ai
Tom’s Hardware, Luke James (reporting the Financial Times): https://www.tomshardware.com/tech-industry/big-tech/big-tech-has-a-tokenmaxxing-habit
AWS Events, re:Invent 2025 keynote: https://www.youtube.com/watch?v=3Y1G9najGiI

Grok Imagine’s 7-Minute AI “Odyssey”: What Developers Should Know

Abhishek Jambla — Mon, 26 Oct 2026 14:50:37 +0000

xAI’s Grok Imagine just posted Episode 1 of an AI-made Odyssey series: about 7 minutes, made with Wonder Studios and director Dev Walker. It arrives three months after Elon Musk promised a full-length, “historically accurate” film by year-end.

How it works (simply)

Image-to-video model: give it a starting image, describe the motion, pick resolution and length.
Model name: grok-imagine-video-1.5, generally available in the xAI API since June 16, 2026.
Short clips: xAI’s own example asks for 10 seconds, so minutes of footage likely means many clips joined together.
Fast mode: 6-second, 720p clips in about 25 seconds (down from 40+).
Sound included: speech, effects and ambience come out in the same pass.

Why it matters

Not a movie yet: this is one episode, and xAI hasn’t shared how it was made.
Mixed quality: TechCrunch says some shots pass for real, but the whole feels flat and lifeless.
Known weak spots: a $100K contest winner showed odd physics, costume changes and a three-eyed Cyclops (TechRadar).
Takeaway: great for short, controlled clips. Story and consistency still need humans.
// Detect dark theme var iframe = document.getElementById('tweet-2113712964419063911-478'); if (document.body.className.includes('dark-theme')) { iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2113712964419063911&theme=dark" }

Media: Episode 1 video · xAI behind-the-scenes poster

References & Credits

TechCrunch – Anthony Ha
Variety – Grok’s first AI-generated “Odyssey” episode
xAI – Grok Imagine Video 1.5 (poster image credit: xAI / Heavy Pulp)
TechRadar – David Nield
Grok Imagine (@imagine) on X

Tech News

Abhishek Jambla — Sun, 25 Oct 2026 13:43:43 +0000

Abhishek Jambla

Oct 25

Why 42% of AI Projects Fail: Lessons from the 2026 AI Graveyard

#agents #ai #aigraveyard #generativeai

2 min read

Why 42% of AI Projects Fail: Lessons from the 2026 AI Graveyard

Abhishek Jambla — Sun, 25 Oct 2026 13:40:10 +0000

TechCrunch’s tracker and S&P Global data reveal that 42% of corporate AI projects are abandoned before reaching production. Even industry leaders and well-funded startups are shutting down flagship tools.

What is breaking—and why it matters for developers:

Platforms Absorb Standalone Tools: OpenAI retired its ChatGPT Atlas browser after 10 months and folded Operator directly into ChatGPT. Startups like Relay and Huxe closed once big platforms added the same automation natively.

Background Agents Replace Separate Apps: Notion Mail shut down because users preferred connecting Notion Custom Agents to automate their existing Gmail or Outlook inboxes instead of switching email clients.

Crushing Server Costs: OpenAI discontinued its Sora video app and API due to heavy compute burn. Meanwhile, Character.ai—despite cutting server costs 33x via custom memory caching—filed for Chapter 11 bankruptcy with a $20M licensing bid from Disney.

Privacy & Overpromising: Microsoft delayed Recall for nearly a year to isolate screenshots inside a secure memory enclave, while Apple’s delayed Siri AI resulted in a $250M settlement.

Takeaway: Skip single-feature wrappers; build integrated workflows with sustainable server costs.

Attached Media:

📊 Architecture Diagram 1: Microsoft Recall VBS Enclave Security Architecture (Windows Experience Blog) (Block diagram showing how Microsoft isolated screenshot decryption inside a hardware-protected memory enclave).

📊 Architecture Diagram 2: Character.AI Cross-Layer KV-Cache & Inference Architecture (Character.AI Engineering Blog) (Diagram illustrating how Character.AI reduced GPU memory footprints by 20x).

🎥 Official Video: Introducing ChatGPT Atlas (OpenAI Official YouTube) (Official launch demo of OpenAI's now-retired AI browser).

References & Credits:

The AI graveyard: A running list of projects and startups that didn't make it — TechCrunch

Generative AI experiences rapid adoption, but with mixed outcomes — S&P Global Market Intelligence

What to know about the Sora discontinuation — OpenAI & ChatGPT Atlas retirement — OpenAI

An Update on Character.ai & Optimizing AI Inference at Character.AI — Character.AI Team

Update on Recall security and privacy architecture — David Weston / Microsoft

Connect Mail to Custom Agents — Notion Documentation

Gemini 4 Argon is Here: 1M Output Tokens, Overnight Code Refactors, and Real-World Speedups

Abhishek Jambla — Sun, 25 Oct 2026 13:05:05 +0000

Google just released gemini-4-argon in Google AI Studio, Vertex AI, and Antigravity. Instead of just generating code snippets, Argon is built to run multi-step engineering workflows from start to finish without losing the plot.

What developers need to know:

Massive 1M Token Output: It features a unified 1,048,576-token context window (text, image, audio, video, PDF) and supports long decode up to 1M output tokens in a single response.

Battle-Tested on Giant Codebases: Google used Argon to untangle a 6-million-line legacy folder inside Chromium’s 40M-line repo in weeks—eliminating an entire class of tab-moving browser crashes.

Autonomous Overnight Optimization: Given a one-paragraph goal, Argon worked overnight on google/adk-python to cut in-memory session prep time by 40% and cold imports by 21%.

Built-In Cyber Defense: Ties for #1 on CWE-Bench v1 (68%) for patching vulnerabilities and cuts indirect prompt injection success down to 0.7%.

Intro Pricing: $2 / 1M input tokens and $10 / 1M output tokens (95% off cached inputs) through November 30, 2026.

What is the biggest legacy refactor on your backlog that you would hand off to an autonomous agent right now?

Gemini 4 Argon: Built to solve your hardest problems by Koray Kavukcuoglu / Google – https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon/

Gemini 4 Argon available on Google Cloud, built to help solve your business' hardest problems by Michael Gerstenhaber / Google Cloud – https://cloud.google.com/blog/products/ai-machine-learning/gemini-4-argon-available-on-google-cloud