Weekly AI Updates (What's New Wednesday)·March 2026·6:10:35

Weekly AI Updates — March 2026 digest (4 episodes)

Karan Rana Host, Weekly AI Updates · Simran Anand Host, Weekly AI Updates · Nitya Host, Weekly AI Updates

The short version

Anthropic/Claude faced public scrutiny (DoD deployment refusal, a Mexican government data-breach allegation, and distillation-attack claims against DeepSeek/Moonshot AI/MiniMax) while simultaneously expanding integrations via Gamma, free-tier Claude Connectors, and the new Claude Code Security scanner.
Google's Nano Banana 2 became the default Gemini image model, adding aspect-ratio syntax, 100+ language text rendering, 4K export, and up to five-character consistency — quickly put to work in marketing tools like Formerly (as heard) and Adobe Firefly's AI-comic workflow.
Agentic infrastructure matured beyond chat: Razorpay's Agent Studio (built on Anthropic's Claude Agent SDK) automates fintech disputes/collections, MCP was explained as the connector layer behind tool-calling, and local-first runtimes (Copa, OpenClaw via Ollama) emerged for privacy-conscious builders.
Computer-use agents advanced quickly: OpenAI's GPT-5.4 scored 75% on the OSWorld benchmark (above the 72% human baseline), Perplexity Computer added desktop/mobile task sync, and Claude doubled its off-peak usage limits for all plans.
Big platforms pushed toward ambient, ecosystem-wide AI: Google signaled a "personal intelligence" strategy weaving Gemini through Maps, Drive, Gmail and Calendar; Perplexity became a system-level assistant on Samsung's Galaxy S26; ChatGPT and Perplexity both teased health-specific assistants.
Copyright and workflow questions recurred around AI-generated creative assets: Adobe stated text-to-image outputs carry full commercial rights, but uploaded reference images require the user's own usage rights — a live tension raised for Formerly's product photoshoots and Firefly's AI comics.
Tools sighted this month: Claude / Claude Code Security / Claude Agent SDK / Claude Connectors; Gemini / Nano Banana 2 / Google Flow; Formerly (as heard); Gamma; Zapier; Perplexity / Perplexity Computer; OpenAI GPT-5.4 / Sora 2; Replit (Replit Animation); Hey Pocket; Razorpay Agent Studio; Google Maps Ask Maps; Google GroundSource / FloodHub; Copa; OpenClaw; Ollama; OpenBrand.sh; FireCrawl; Jina AI; Adobe Firefly; Canva; Midjourney v8; Microsoft MAI-Image-1; Sarvam AI; HeyGen; Copy.ai; Unitree; DeepSeek.

The concepts

Image & Video Generation Levels Up: Nano Banana 2, Sora 2, and AI Comics

Google's Nano Banana 2 became the default Gemini image model this month, adding production-ready aspect-ratio control, 100+ language text rendering, 4K export, and up to five consistent characters across a generated sequence. Rivals moved in parallel — OpenAI's Sora 2 added custom characters, 20-second clips, and scene continuation, while Adobe Firefly's style/composition reference system was pitched for AI comics and storyboards. Marketing tools (Formerly, Firefly) still lack automated layout/arrangement, so creators stitch single-image generations into sequences by hand.

Nano Banana 2 added explicit aspect-ratio syntax (e.g. "AR 1:1", "AR 9:16") so one prompt can be resized for Instagram, GDN banners, or skyscraper ads without a separate crop step (0:39:32, l3535997).

Multilingual text rendering (100+ languages) and character identity/outfit consistency (up to 5 characters) were demoed live in Hindi, Greek, and Arabic with roughly 95-99% spelling accuracy per informal room testing (0:43:33, l3535997).

Google Labs' Formerly (as heard) launched in India: a "business DNA" step (logo/fonts/colors/tagline) feeds AI-generated product photoshoots and ad creatives, but has no ad-platform export/integration yet (0:44:06, l3535996).

OpenAI's Sora 2 added custom-character video generation, 20-second clips, batch generation, and scene continuation via its developer API (0:14:41, l3535999).

Adobe Firefly's "use as style reference" and "use as composition reference" controls lock lighting/camera-angle/mood across a comic panel sequence; Adobe states text-to-image carries full commercial copyright, but uploaded reference images require the user's own usage rights (0:58:24, l3536002).

Agents Go Infrastructural: MCP, Claude Agent SDK, and Local-First Runtimes

This month's sessions treated agents as backend infrastructure rather than chatbots: Razorpay shipped a fintech Agent Studio built on Anthropic's Claude Agent SDK with sub-agents for chargeback disputes and cart-abandonment recovery, MCP was explained as the connector standard letting apps like a brokerage or Zerodha expose data to any LLM client, and open-source local-agent runtimes (Copa, OpenClaw, both as heard) running through Ollama were positioned for privacy-sensitive use cases. Hosts repeatedly framed human-in-the-loop review as the default for external-facing agent actions.

Razorpay's Agent Studio (built on Anthropic's Claude Agent SDK) uses a master agent that decomposes financial-ops tasks — chargebacks, subscription recovery, invoicing — into specialized sub-agents (0:19:57, l3535999).

MCP was explained via a brokerage-app example: once a platform exposes itself via MCP, a user can ask any connected LLM (ChatGPT, Perplexity, Claude) about their portfolio directly (0:39:51, l3535996).

Anthropic's new Claude Code Security scans codebases for vulnerabilities and suggests patches, but keeps a human in the loop rather than auto-deploying fixes (0:21:44, l3535996).

Copa and OpenClaw (both as heard) are open-source AI agent runtimes that pair with Ollama for fully offline, local-model agents with long-term (episodic + semantic) memory (0:51:13, l3535999).

A live audience poll showed a roughly even split on full automation vs. approval gates; consensus leaned toward hands-off for internal/small-team tasks but mandatory human approval before anything external-facing goes live (0:23:44, l3535996).

AI Goes Ambient: Ecosystem Connectors, On-Device Push, and Cross-Device Agents

Big platforms pushed AI further into daily ambient use this month: Claude opened its 150+ integrations to free-tier users, Perplexity became a system-level assistant on Samsung's Galaxy S26 ("Hey Plex") and added cross-device task sync for Perplexity Computer, and Google signaled a broader "personal intelligence" push weaving Gemini through Maps, Drive, Gmail, and Calendar. The throughline across episodes: fewer app switches, with one assistant orchestrating many connected tools via MCP-style connectors.

Claude Connectors (MCP-based integrations — Gmail, Sheets, Notion, Airtable, Drive) expanded to free-tier users, not just paid plans (0:35:51, l3535996).

Perplexity partnered with Samsung to embed itself as a system-level, wake-word ("Hey Plex") assistant in the upcoming Galaxy S26 line (0:35:51, l3535996).

Perplexity Computer added task syncing across desktop and mobile so an in-progress agent task can be resumed on a phone (0:28:56, l3535999).

Google Maps' new "Ask Maps" feature uses Gemini to answer conversational questions about a place (reviews, itineraries, EV-charging-area recommendations), live in the US with an India rollout planned (0:10:40, l3535999).

Karan framed Google's Gemini-across-every-app strategy as "personal intelligence," contrasting it with Microsoft's B2B-first Azure approach and Apple's later entry into the space (0:13:53, l3536002).

Every concept, three clicks deep

The same concepts as a quick reference: the closed row is the glance, open is the study card, and every timestamp jumps into the recording.

01Image & Video Generation Levels Up: Nano Banana 2, Sora 2, and AI ComicsGoogle's Nano Banana 2 became the default Gemini image model this month, adding production-ready aspect-rat…›

OpenAI's Sora 2 added custom-character video generation, 20-second clips, batch generation, and scene continuation via its developer API (0:14:41, l3535999).

02Agents Go Infrastructural: MCP, Claude Agent SDK, and Local-First RuntimesThis month's sessions treated agents as backend infrastructure rather than chatbots: Razorpay shipped a fin…›

Anthropic's new Claude Code Security scans codebases for vulnerabilities and suggests patches, but keeps a human in the loop rather than auto-deploying fixes (0:21:44, l3535996).

Copa and OpenClaw (both as heard) are open-source AI agent runtimes that pair with Ollama for fully offline, local-model agents with long-term (episodic + semantic) memory (0:51:13, l3535999).

03AI Goes Ambient: Ecosystem Connectors, On-Device Push, and Cross-Device AgentsBig platforms pushed AI further into daily ambient use this month: Claude opened its 150+ integrations to f…›

Claude Connectors (MCP-based integrations — Gmail, Sheets, Notion, Airtable, Drive) expanded to free-tier users, not just paid plans (0:35:51, l3535996).

Perplexity partnered with Samsung to embed itself as a system-level, wake-word ("Hey Plex") assistant in the upcoming Galaxy S26 line (0:35:51, l3535996).

Perplexity Computer added task syncing across desktop and mobile so an in-progress agent task can be resumed on a phone (0:28:56, l3535999).

Karan framed Google's Gemini-across-every-app strategy as "personal intelligence," contrasting it with Microsoft's B2B-first Azure approach and Apple's later entry into the space (0:13:53, l3536002).

Tools referenced

Tool	Coverage	Moment	Context

Action items

Extraction notes

This page was built from an auto-generated transcript, which garbles product and people's names. Those were corrected silently in everything above and logged here for transparency. The warnings flag claims that were true on the recording day but change fast.

Transcript corrections applied

The transcript says	The trainer actually means
Formerly / Parmalee / family / Pameli	Google Labs' new AI marketing/social-campaign tool (business-DNA + photoshoot features) demoed in the Mar 4 episode — exact product name uncertain as heard, transcribed inconsistently across the session.
Copa / Coper / Cobo	The open-source AI agent framework with long-term memory that pairs with Ollama, discussed in the Mar 18 episode — spelling inconsistent in the transcript.
OpenClaw	A local-agent runtime discussed alongside Ollama as an official LLM provider integration — exact product name uncertain as heard.
GroundSource	Google Research's synthetic flood-forecasting dataset/model (2.6M mapped flood events, results published via FloodHub) — may also be styled "Ground Source."
Sarvam / Servam	Sarvam AI, the India-focused sovereign AI startup that reportedly secured Nvidia funding, discussed across the Mar 4 and Mar 25 episodes.
Hey Pocket	"Pocket," the offline AI meeting recorder hardware (magnetic phone attachment, 64GB storage, speaker diarization) demoed in the Mar 11 episode.

True on recording day — verify before relying

T
h
i
s
d
i
g
e
s
t
s
u
m
m
a
r
i
z
e
s
i
n
f
o
r
m
a
l
,
h
o
s
t
-
n
a
r
r
a
t
e
d
n
e
w
s
r
o
u
n
d
u
p
s
r
e
c
o
r
d
e
d
l
i
v
e
b
e
t
w
e
e
n
M
a
r
c
h
4
a
n
d
M
a
r
c
h
2
5
,
2
0
2
6
.
P
r
o
d
u
c
t
n
a
m
e
s
,
b
e
n
c
h
m
a
r
k
f
i
g
u
r
e
s
,
f
u
n
d
i
n
g
a
m
o
u
n
t
s
,
a
n
d
f
e
a
t
u
r
e
c
l
a
i
m
s
(
e
.
g
.
N
a
n
o
B
a
n
a
n
a
2
'
s
l
a
n
g
u
a
g
e
/
r
e
s
o
l
u
t
i
o
n
s
p
e
c
s
,
G
P
T
-
5
.
4
'
s
O
S
W
o
r
l
d
s
c
o
r
e
,
S
a
r
v
a
m
A
I
'
s
N
v
i
d
i
a
f
u
n
d
i
n
g
,
C
l
a
u
d
e
'
s
o
f
f
-
p
e
a
k
u
s
a
g
e
w
i
n
d
o
w
)
a
r
e
a
s
s
t
a
t
e
d
b
y
t
h
e
h
o
s
t
s
o
n
a
i
r
a
n
d
h
a
v
e
n
o
t
b
e
e
n
i
n
d
e
p
e
n
d
e
n
t
l
y
v
e
r
i
f
i
e
d
a
g
a
i
n
s
t
v
e
n
d
o
r
d
o
c
u
m
e
n
t
a
t
i
o
n
.
S
e
v
e
r
a
l
p
r
o
d
u
c
t
n
a
m
e
s
w
e
r
e
u
n
c
l
e
a
r
o
r
i
n
c
o
n
s
i
s
t
e
n
t
l
y
p
r
o
n
o
u
n
c
e
d
i
n
t
h
e
a
u
d
i
o
(
s
e
e
a
s
r
C
o
r
r
e
c
t
i
o
n
s
)
a
n
d
s
h
o
u
l
d
b
e
c
o
n
f
i
r
m
e
d
b
e
f
o
r
e
b
e
i
n
g
c
i
t
e
d
a
u
t
h
o
r
i
t
a
t
i
v
e
l
y
.
G
i
v
e
n
t
h
e
p
a
c
e
o
f
A
I
p
r
o
d
u
c
t
r
e
l
e
a
s
e
s
,
f
e
a
t
u
r
e
s
e
t
s
,
p
r
i
c
i
n
g
,
a
n
d
a
v
a
i
l
a
b
i
l
i
t
y
d
e
s
c
r
i
b
e
d
h
e
r
e
m
a
y
a
l
r
e
a
d
y
b
e
o
u
t
d
a
t
e
d
b
y
t
h
e
t
i
m
e
t
h
i
s
d
i
g
e
s
t
i
s
r
e
a
d
—
t
r
e
a
t
a
l
l
s
p
e
c
i
f
i
c
s
a
s
a
s
n
a
p
s
h
o
t
o
f
w
h
a
t
w
a
s
r
e
p
o
r
t
e
d
i
n
M
a
r
c
h
2
0
2
6
,
n
o
t
a
s
c
u
r
r
e
n
t
f
a
c
t
.