Weekly AI Updates (What's New Wednesday)·November 2025·7:20:29

Weekly AI Updates — November 2025 digest (4 episodes)

Karan Rana (as heard) Host, Weekly AI Updates · Dilip (as heard, guest mentor, Nov 26 episode) Guest mentor, Weekly AI Updates

▶local recording · LMS lesson

The short version

Google shipped Gemini 3 plus a new Google Antigravity dev platform and 'GenUI' (interactive, app-like answers instead of text walls), rolling out to the US first.
Big-tech AI rentals continued: Apple will pay Google roughly $1B/year for a privately-hosted Gemini to power Siri, echoing Microsoft's earlier ChatGPT-rental phase before it built its own MAI models.
Agentic browsers went mainstream — ChatGPT Atlas can log in and act on a user's behalf (fill forms, edit Slides); Perplexity's Comet went free-for-all and drew an Amazon lawsuit over scraping and buying without human page visits.
Voice AI shifted from cloning free-for-all to licensing: ElevenLabs launched an 'Iconic' marketplace for celebrity voices amid Indian lawsuits (Arijit Singh, Asha Bhosle, Amitabh Bachchan) over unauthorized clones; Cartesia Sonic 3 and MiniMax Speech 2.6 emerged as low-latency rivals.
India-specific AI investment accelerated: free Gemini via Jio (425M users), IIT Madras's Indic LLM Arena (with Google Cloud), and OpenAI's INDQA benchmark (12 languages, 2,000+ expert questions).
AI video tooling got a hands-on deep dive (OpusClip vs Gling clipping/virality-score workflows, Hailuo 2.3 vs Google Veo 3.1) before the month closed with fundamentals: Claude Opus 4.5 broke the 80% coding-benchmark mark, and a guest mentor ran a nontechnical fine-tuning-vs-RAG primer.
Tools sighted this month: Gemini 3/Enterprise, Google Antigravity, ChatGPT Go/Atlas, OpenAI Aardvark, Google Pomelli, Unitree G1, Hailuo 2.3, Veo 3.1, NotebookLM, Cartesia Sonic 3, MiniMax Speech 2.6, Morphic, Manus 1.5, PayPal-in-ChatGPT, ElevenLabs Iconic, Perplexity Comet, OpusClip, Gling, Claude Opus 4.5, Nano Banana Pro, Kimi K2 Thinking, DeepSeek OCR, Seedream, Nebius Token Factory, Together AI, Unsloth.

The concepts

Agentic AI browsers take actions online

ChatGPT Atlas and Perplexity Comet moved in November from Q&A copilots to agents that log into a user's own accounts and complete tasks — filling forms, editing slide decks, or buying items — via a toggled logged-in or logged-out mode. The shift triggered a legal fight when Amazon sued Perplexity's Comet agent for scraping and purchasing on Amazon without human page visits, since Amazon's ad revenue depends on human traffic.

Atlas shifts from answering questions to completing web tasks — filling forms, fetching info (0:29:50, l3029792)

Logged-in mode lets Atlas act across a user's authenticated sites; Sam Altman flagged the risk of third-party attackers exploiting it (0:31:51, l3029792)

Live demo: Atlas agent mode autonomously inserts a new slide at the top of a Google Slides deck (0:33:52, l3029792)

Perplexity's agent-based browsing goes free for everyone, facing early pushback from Amazon on shopping automation (0:39:55, l3029792)

Voice AI moves from cloning to licensing

ElevenLabs launched an 'Iconic' marketplace letting public figures and estates approve paid, contracted use of their voice, a direct response to Indian lawsuits over unauthorized voice cloning. New low-latency rivals Cartesia Sonic 3 and MiniMax Speech 2.6 target the call-center/voice-agent market ElevenLabs has long dominated.

Backdrop: Arijit Singh's voice cloned without consent for music sparks the discussion (0:24:32, l3075862)

Asha Bhosle and Amitabh Bachchan file suits in Mumbai High Court over unauthorized AI voice cloning, citing personality rights (0:26:33, l3075862)

ElevenLabs launches its Iconic marketplace, licensing 28+ celebrity voices (e.g. Michael Caine) for brand campaigns, podcasts, and audiobooks (0:28:37, l3075862)

Licensing flow: requester submits project scope; the voice's rights holder or estate approves or denies each use case before a contract is signed (0:34:42, l3075862)

Fine-tuning vs RAG, a nontechnical primer

A guest mentor (as heard, 'Dilip') distinguished RAG — likened to a stretchable playbook a trained model consults per query — from fine-tuning, a permanent 'mold' that retrains the model itself, using a soccer-player-to-rugby analogy. The session then demoed building a LoRA fine-tune of GPT-4.1 nano in OpenAI Playground from a creator's own LinkedIn posts.

RAG = LLM + memory (a playbook consulted per query); fine-tuning = retraining the model itself, a permanent change (0:50:53, l3082672)

LoRA (low-rank adaptation) modifies only a small part of the model cheaply vs full training, which 'unlearns and relearns' at far higher cost (0:58:55, l3082672)

Live demo: convert ~25-30 LinkedIn posts into a JSONL prompt/output dataset and fine-tune GPT-4.1 nano in OpenAI Playground for about $1 and ~30 minutes (1:11:11, l3082672)

Nebius AI's Token Factory offers one-click fine-tuning/deployment for open-source models; Together AI and Unsloth recommended for code-based fine-tuning (1:31:34, l3082672)

Rule of thumb: fine-tune for a fixed persistent personality/tone; use RAG when citations or frequently updated knowledge are required (1:39:44, l3082672)

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.

01Agentic AI browsers take actions onlineChatGPT Atlas and Perplexity Comet moved in November from Q&A copilots to agents that log into a user's own…›

Atlas shifts from answering questions to completing web tasks — filling forms, fetching info (0:29:50, l3029792)

Logged-in mode lets Atlas act across a user's authenticated sites; Sam Altman flagged the risk of third-party attackers exploiting it (0:31:51, l3029792)

Live demo: Atlas agent mode autonomously inserts a new slide at the top of a Google Slides deck (0:33:52, l3029792)

Perplexity's agent-based browsing goes free for everyone, facing early pushback from Amazon on shopping automation (0:39:55, l3029792)

02Voice AI moves from cloning to licensingElevenLabs launched an 'Iconic' marketplace letting public figures and estates approve paid, contracted use…›

Backdrop: Arijit Singh's voice cloned without consent for music sparks the discussion (0:24:32, l3075862)

Asha Bhosle and Amitabh Bachchan file suits in Mumbai High Court over unauthorized AI voice cloning, citing personality rights (0:26:33, l3075862)

ElevenLabs launches its Iconic marketplace, licensing 28+ celebrity voices (e.g. Michael Caine) for brand campaigns, podcasts, and audiobooks (0:28:37, l3075862)

Licensing flow: requester submits project scope; the voice's rights holder or estate approves or denies each use case before a contract is signed (0:34:42, l3075862)

03Fine-tuning vs RAG, a nontechnical primerA guest mentor (as heard, 'Dilip') distinguished RAG — likened to a stretchable playbook a trained model co…›

RAG = LLM + memory (a playbook consulted per query); fine-tuning = retraining the model itself, a permanent change (0:50:53, l3082672)

LoRA (low-rank adaptation) modifies only a small part of the model cheaply vs full training, which 'unlearns and relearns' at far higher cost (0:58:55, l3082672)

Live demo: convert ~25-30 LinkedIn posts into a JSONL prompt/output dataset and fine-tune GPT-4.1 nano in OpenAI Playground for about $1 and ~30 minutes (1:11:11, l3082672)

Nebius AI's Token Factory offers one-click fine-tuning/deployment for open-source models; Together AI and Unsloth recommended for code-based fine-tuning (1:31:34, l3082672)

Rule of thumb: fine-tune for a fixed persistent personality/tone; use RAG when citations or frequently updated knowledge are required (1:39:44, l3082672)

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
Aardvark / Arduard	OpenAI Aardvark (AI code-security agent)
Pomoly / Pomeli	Google Pomelli
Helio / Hilu	Hailuo 2.3 (MiniMax video model)
Cartesian (company)	Cartesia; its TTS model is Sonic 3
Kine k2 / Kimi Kato thinking	Kimi K2 Thinking (Moonshot AI)
Nabeas / Maybes	Nebius AI (Token Factory)

True on recording day — verify before relying

T
h
i
s
i
s
a
r
o
l
l
i
n
g
A
I
-
n
e
w
s
r
o
u
n
d
u
p
s
n
a
p
s
h
o
t
f
r
o
m
N
o
v
e
m
b
e
r
2
0
2
5
(
4
e
p
i
s
o
d
e
s
)
;
t
o
o
l
n
a
m
e
s
,
p
r
i
c
i
n
g
,
f
e
a
t
u
r
e
a
v
a
i
l
a
b
i
l
i
t
y
(
G
e
m
i
n
i
3
U
S
-
o
n
l
y
r
o
l
l
o
u
t
,
C
h
a
t
G
P
T
G
o
p
r
i
c
i
n
g
,
P
o
m
e
l
l
i
'
s
r
e
g
i
o
n
a
l
b
e
t
a
)
,
a
n
d
l
e
g
a
l
d
i
s
p
u
t
e
s
(
A
m
a
z
o
n
v
.
P
e
r
p
l
e
x
i
t
y
)
m
a
y
h
a
v
e
c
h
a
n
g
e
d
m
a
t
e
r
i
a
l
l
y
b
y
t
h
e
t
i
m
e
t
h
i
s
i
s
r
e
a
d
—
v
e
r
i
f
y
c
u
r
r
e
n
t
s
t
a
t
u
s
b
e
f
o
r
e
c
i
t
i
n
g
a
s
f
a
c
t
.