Weekly AI Updates (What's New Wednesday)·July 2025·3:37:27

Weekly AI Updates — July 2025 digest (2 episodes)

Karan Rana Host, Weekly AI Updates (Senior Product Marketing Manager, Swiggy) · Bhardwaj / Micah (as-heard, session moderators) Co-host, session moderator / Slido Q&A wrangler

▶local recording · LMS lesson

The short version

ChatGPT agent (merging Operator + Deep Research) rolled out July 17 to Pro/Plus/Team users; Sam Altman warned against trusting it with banking or financial tasks (as-heard).
Google shipped a cluster of agentic search features: Gemini-powered AI mode 'power calling' (books appointments with low-stake merchants like dry cleaners), Web Guide (thematic SERP clustering, US rollout), and Chrome 'Store Summaries'; Microsoft answered with Edge 'Copilot Mode.'
Video generation jumped forward: Google VO3 + Google Flow added lip-synced audio and scene-extend, with SynthID invisible watermarks for provenance; Grok 4 (xAI) launched with a larger context window and controversial 3D 'Companion' avatars (Ani, Rudy); Google Imagen 4 was praised for finally fixing hands and on-image text.
Genspark's 'super agent' (as-heard, Chinese startup, reportedly raised ~$100M at a >$500M valuation) became the presenter's new go-to over ChatGPT agent — one prompt fans out across ~80 tools / 9 LLMs / 18 video models to produce podcasts, video, images, websites, decks, sheets, and fact-checked 'Spark Pages,' with free daily credits.
Sam Altman disclosed that ChatGPT chats are not legally privileged and could be produced as court evidence even if deleted, applying to free and Pro tiers alike; presenter urged disabling chat history/personalization and never pasting company or confidential data into any AI tool.
GPT-5 was teased by the presenter for an 'early-to-mid August' launch — explicitly flagged as speculation, not confirmed on-air.
Macro framing repeated across both episodes: Google/Microsoft/Meta/Amazon combined ~$300B AI investment in 2025; Sam Altman flagged customer-service roles at risk; TCS layoffs (thousands, cause officially unstated) cited as a real-world signal.
Tools sighted this month: ChatGPT agent, Google AI mode power calling, Web Guide, Chrome Store Summaries, Edge Copilot Mode, VO3 / VO3 Fast, Google Flow, SynthID, Imagen 4, Grok 4 + Companions, Genspark (AI Pods, Spark Pages), Runner H, Manus, n8n, Perplexity Comet, NotebookLM, HeyGen, Flux, Gamma, Napkin.ai, Runway, Kling v1.6 Pro, DeepSeek R1, Supabase.

The concepts

Theme: the agent-and-browser land grab (ChatGPT agent, Google Web Guide/power calling, Genspark super agent)

Across both episodes the presenter tracked a shift from single-purpose chatbots to multi-tool 'super agents' and AI-native browsing layers. OpenAI, Google, Microsoft, and the newer entrant Genspark all shipped agentic features in the same two-week window, each racing to own the 'do it for me' layer of the browser and OS.

ChatGPT agent launched July 17, merging Operator and Deep Research; available to Pro/Plus/Team tiers; Altman cautioned against giving it banking/financial access (0:08:09, l2869746)

Google AI mode 'power calling' (Gemini-powered) can call low-stake merchants (dry cleaners, pet groomers) on a user's behalf, announcing itself as a Google bot; merchants can opt out (0:18:15, l2869746)

Jul 30: Google Web Guide clusters SERP results thematically (US rollout); Chrome added 'Store Summaries'; Edge added 'Copilot Mode' to compare open tabs (0:18:51, l2871239)

Genspark 'super agent' (as-heard, Chinese company) consolidates ~80 tools / 9 LLMs / 18 video models into one chat window — AI Pods, video/image gen, no-code website builder, one-prompt multi-deliverable runs, and a built-in 'fact check' pill on slides (0:29:00, l2871239)

Presenter's comparison: Genspark's deep-research output ('Spark Pages') was better formatted and navigable than ChatGPT agent's bullet-heavy output on an identical prompt (0:55:20, l2871239)

Theme: the video/image model arms race (VO3, Grok 4, Imagen 4)

July saw a burst of multimodal generation upgrades, especially in video: audio finally got matched to lip-synced video generation, context windows grew, and long-standing image artifacts (hands, on-image text) were reported fixed. Presenter treated this as a snapshot of 'this month's best model,' expected to be obsolete within weeks.

Google VO3 + Google Flow generate lip-synced video with matched sound effects from a single photo; VO3 Fast (intro tier) limited to 3 videos/day, 8 seconds, 720p; 'Extend' chains additional 8-second clips (0:26:19, l2869746)

Grok 4 (xAI) launched with an expanded context window (as-heard ~256k tokens vs Grok 3's ~128k) and a controversial 'Companion' feature — 3D avatar characters (Ani; Rudy, a panda for kids mode) that see and talk via camera (0:52:41, l2869746)

Google Imagen 4 called out as finally fixing garbled fingers/hands and misspelled on-image text (1:08:49, l2869746)

SynthID invisible watermarks are embedded in all Veo-generated video for provenance; YouTube Content ID separately detects reuse of AI-generated clips (1:24:57, l2869746)

Theme: AI chats are not privileged — data privacy and discretion warnings

A recurring caution across the July sessions: unlike lawyer/doctor/accountant conversations, AI chat logs are not legally protected and default settings often retain or reuse conversation data. The presenter repeatedly urged auditing privacy toggles and never pasting confidential or company data into any AI tool.

Sam Altman stated ChatGPT chats are not legally privileged; even deleted chats could be subpoenaed as court evidence, true on both free and Pro tiers (0:22:55, l2871239)

Presenter's own anecdote: ChatGPT retained his wife's mention of his name from 50 days earlier and referenced it unprompted in a later chat — default memory/history retention in action (1:25:59, l2871239)

Genspark's default 'AI data retention' setting reuses user chats/searches for training unless manually toggled off — presenter demonstrated disabling it live (1:30:03, l2871239)

General guidance: never upload real company data, use masked/dummy data with broad prompts, and check each tool's settings for history/personalization toggles (1:28:02, l2871239)

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.

01Theme: the agent-and-browser land grab (ChatGPT agent, Google Web Guide/power calling, Genspark super agent)Across both episodes the presenter tracked a shift from single-purpose chatbots to multi-tool 'super agents…›

ChatGPT agent launched July 17, merging Operator and Deep Research; available to Pro/Plus/Team tiers; Altman cautioned against giving it banking/financial access (0:08:09, l2869746)

Jul 30: Google Web Guide clusters SERP results thematically (US rollout); Chrome added 'Store Summaries'; Edge added 'Copilot Mode' to compare open tabs (0:18:51, l2871239)

Presenter's comparison: Genspark's deep-research output ('Spark Pages') was better formatted and navigable than ChatGPT agent's bullet-heavy output on an identical prompt (0:55:20, l2871239)

02Theme: the video/image model arms race (VO3, Grok 4, Imagen 4)July saw a burst of multimodal generation upgrades, especially in video: audio finally got matched to lip-s…›

Google Imagen 4 called out as finally fixing garbled fingers/hands and misspelled on-image text (1:08:49, l2869746)

SynthID invisible watermarks are embedded in all Veo-generated video for provenance; YouTube Content ID separately detects reuse of AI-generated clips (1:24:57, l2869746)

03Theme: AI chats are not privileged — data privacy and discretion warningsA recurring caution across the July sessions: unlike lawyer/doctor/accountant conversations, AI chat logs a…›

Sam Altman stated ChatGPT chats are not legally privileged; even deleted chats could be subpoenaed as court evidence, true on both free and Pro tiers (0:22:55, l2871239)

Genspark's default 'AI data retention' setting reuses user chats/searches for training unless manually toggled off — presenter demonstrated disabling it live (1:30:03, l2871239)

General guidance: never upload real company data, use masked/dummy data with broad prompts, and check each tool's settings for history/personalization toggles (1:28:02, l2871239)

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
Hagen	HeyGen (as-heard, verify)
Runnerage / Runner Edge / Jetstock	Runner H (as-heard, verify)
Anyten	n8n
Complexity / public cities	Perplexity
Watson audio overview	NotebookLM audio overview (as-heard)
SYNTIDs / synth SYNT IDs	SynthID
Claude Gold	Claude Code (as-heard)
SuperBase	Supabase
it news (fintech stock broking platform)	possibly Zerodha (as-heard, unverified)

True on recording day — verify before relying

T
h
i
s
i
s
a
n
e
w
s
-
d
i
g
e
s
t
s
n
a
p
s
h
o
t
o
f
J
u
l
y
2
0
2
5
A
I
-
t
o
o
l
l
a
u
n
c
h
e
s
(
C
h
a
t
G
P
T
a
g
e
n
t
,
G
e
n
s
p
a
r
k
s
u
p
e
r
a
g
e
n
t
,
V
O
3
/
G
r
o
k
4
/
I
m
a
g
e
n
4
,
G
o
o
g
l
e
W
e
b
G
u
i
d
e
)
;
e
v
e
r
y
v
e
r
s
i
o
n
n
u
m
b
e
r
,
p
r
i
c
e
,
c
r
e
d
i
t
l
i
m
i
t
,
a
n
d
l
a
u
n
c
h
d
a
t
e
i
s
a
s
-
h
e
a
r
d
f
r
o
m
t
h
e
p
r
e
s
e
n
t
e
r
o
n
-
a
i
r
a
n
d
s
e
v
e
r
a
l
c
l
a
i
m
s
(
G
P
T
-
5
t
i
m
i
n
g
,
c
o
n
t
e
x
t
-
w
i
n
d
o
w
s
i
z
e
s
,
c
o
m
p
a
n
y
n
a
m
e
s
)
a
r
e
e
x
p
l
i
c
i
t
l
y
s
p
e
c
u
l
a
t
i
v
e
o
r
m
i
s
-
t
r
a
n
s
c
r
i
b
e
d
—
r
e
w
a
t
c
h
t
h
e
e
p
i
s
o
d
e
o
r
v
e
r
i
f
y
i
n
d
e
p
e
n
d
e
n
t
l
y
b
e
f
o
r
e
t
r
e
a
t
i
n
g
s
p
e
c
i
f
i
c
s
a
s
c
u
r
r
e
n
t
.