Weekly AI Updates (What's New Wednesday)·June 2026·6:12:34

Weekly AI Updates — June 2026 digest (4 episodes)

Amit (as heard, also transcribed "Amit Shivatsa" / "Amish Yodsagoti") Host, Weekly AI Updates · Alejo Host, Weekly AI Updates · Arfa Co-host/moderator (chat and Slido), Weekly AI Updates

▶local recording · LMS lesson

The short version

Anthropic's flagship model launched twice this month — Claude Opus 4.8 (Jun 3) then the more capable Fable 5/Mythos (~Jun 9) — but Fable 5 was banned for non-US users days later after an AWS-reported jailbreak triggered a US government directive; Anthropic is now proposing government-ID verification (from Jul 8) to restore access.
Chinese open-weight models kept closing the cost/quality gap all month — Alibaba's Qwen-based model (~1/3 the price of Opus 4.8) and z.ai's GLM 5.2 were shown matching or beating Claude 4.7/4.8 on several benchmarks, and GLM 5.2 can be wired directly into the Claude Code harness.
Huge compute/money moves: SpaceX's roughly $2T IPO (raised $75B, Elon Musk became the first trillionaire), a Google-to-SpaceX $1B/month data-center compute deal, and SpaceX's reported $60B acquisition of Cursor all raised open questions about an AI valuation bubble.
AI "assistants" kept turning into standing agents rather than one-off chat tools — Google's Gemini Spark, ChatGPT scheduled tasks, an increasingly agentic NotebookLM, and Claude/Codex "record and replay" computer-use automation all let users hand off a recurring goal instead of a single prompt.
Physical-world AI accelerated: Boston Dynamics' Atlas learned to lift an unseen fridge after training almost entirely in simulation, NVIDIA released an open physical-AI foundation model and partnered with Boston Dynamics on factory robots, and Midjourney showed a low-cost, near-real-time body-imaging device.
Office and creative suites consolidated around a single AI layer: Microsoft put Copilot (with new "Work IQ" adaptive reasoning) and Claude into Excel/Word/PowerPoint, Adobe unified Firefly into one chat assistant across Photoshop/Illustrator/InDesign, and Lovable/Canva/ChatGPT added click-to-edit visual AI editing.
Tools sighted this month: OpenAI ChatGPT/Codex; Anthropic Claude (Opus 4.8, Fable 5/Mythos), Claude Code, Claude for Excel/Word/PowerPoint; Alibaba Qwen; z.ai GLM 5.2; Google Gemini, Gemini Spark, Gemma 4, NotebookLM, Google Vids; NVIDIA physical-AI model; Boston Dynamics Atlas; Microsoft Copilot/Work IQ; Cursor; Perplexity; ElevenLabs; Adobe Firefly; Midjourney; Kimi; Coinbase AI trading; Salesforce Agentforce/Fin; Figma; HyperFrames; Canva Magic Layers; Lovable; Apple Siri; Notion; SpaceX/Starlink.

The concepts

Assistants becoming standing agents

Across June, major AI platforms shifted from one-off question-and-answer toward standing agentic behavior: Gemini Spark, ChatGPT scheduled tasks, a more self-directed NotebookLM, and Claude/Codex computer-use "record and replay" all let a user hand over a recurring goal instead of a single prompt. Presenters framed this as AI moving from "answer me now" to "keep watch for me."

Google's Gemini Spark keeps persistent context across Gmail and Drive, but launched only for US Google AI Ultra subscribers (0:32:57, l3963414)

ChatGPT scheduled tasks let a user set a recurring prompt (e.g., a daily AI-news brief) that runs and notifies automatically (0:37:26, l3963420)

NotebookLM can now find its own sources, run code, and generate diagrams instead of just summarizing uploaded documents (0:30:53, l3963417)

Claude and Codex added record-and-replay style computer-use automation that watches a workflow once and repeats it (1:03:47, l3963420)

Open-weight models closing the gap

Chinese open-weight releases — Alibaba's Qwen-based model and z.ai's GLM 5.2 — were shown matching or beating Claude Opus 4.7/4.8 on several public benchmarks at a fraction of the price, and GLM 5.2 could be pointed directly at the Claude Code harness. Hosts argued frontier capability is no longer a US-only moat and that workflows should be built model-agnostic.

Alibaba's Qwen-based model was priced at roughly one-third of Claude Opus 4.8 in the comparison shown (0:26:49, l3963414)

z.ai's GLM 5.2 (open, ~1M context) beat Claude 4.7 and matched some 4.8 results on benchmarks shown (0:49:34, l3963420)

GLM 5.2 can be swapped into the Claude Code harness via ANTHROPIC_AUTH_TOKEN/base-URL environment variables pointed at z.ai's API (0:18:44, l3963417)

Anthropic's Dario Amodei reportedly said other open-model companies could reach Mythos-level quality within 6-12 months (0:20:44, l3963417)

Physical AI moving from lab demo to factory floor

NVIDIA, Boston Dynamics, and others pushed multimodal "physical AI" toward commercial deployment this month: robots trained almost entirely in simulation before real-world use, a new NVIDIA-Boston Dynamics factory partnership, and reports of camera-headband data collection from human workers to train humanoid-robot dexterity.

Boston Dynamics' Atlas relocated a refrigerator it had never seen after training purely in simulation, not physical trial-and-error (0:53:18, l3963414)

NVIDIA released an open multimodal foundation model for robots that processes text, image, video, audio, and physical actions together (0:22:25, l3963416)

NVIDIA announced a partnership with Boston Dynamics to push these robot models into factories (0:59:19, l3963417)

Reports described workers paid to wear camera headbands while cleaning, stitching, and folding to train humanoid-robot dexterity models (1:07:49, l3963420)

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.

01Assistants becoming standing agentsAcross June, major AI platforms shifted from one-off question-and-answer toward standing agentic behavior:…›

Google's Gemini Spark keeps persistent context across Gmail and Drive, but launched only for US Google AI Ultra subscribers (0:32:57, l3963414)

ChatGPT scheduled tasks let a user set a recurring prompt (e.g., a daily AI-news brief) that runs and notifies automatically (0:37:26, l3963420)

NotebookLM can now find its own sources, run code, and generate diagrams instead of just summarizing uploaded documents (0:30:53, l3963417)

Claude and Codex added record-and-replay style computer-use automation that watches a workflow once and repeats it (1:03:47, l3963420)

02Open-weight models closing the gapChinese open-weight releases — Alibaba's Qwen-based model and z.ai's GLM 5.2 — were shown matching or beati…›

Alibaba's Qwen-based model was priced at roughly one-third of Claude Opus 4.8 in the comparison shown (0:26:49, l3963414)

z.ai's GLM 5.2 (open, ~1M context) beat Claude 4.7 and matched some 4.8 results on benchmarks shown (0:49:34, l3963420)

GLM 5.2 can be swapped into the Claude Code harness via ANTHROPIC_AUTH_TOKEN/base-URL environment variables pointed at z.ai's API (0:18:44, l3963417)

Anthropic's Dario Amodei reportedly said other open-model companies could reach Mythos-level quality within 6-12 months (0:20:44, l3963417)

03Physical AI moving from lab demo to factory floorNVIDIA, Boston Dynamics, and others pushed multimodal "physical AI" toward commercial deployment this month…›

Boston Dynamics' Atlas relocated a refrigerator it had never seen after training purely in simulation, not physical trial-and-error (0:53:18, l3963414)

NVIDIA released an open multimodal foundation model for robots that processes text, image, video, audio, and physical actions together (0:22:25, l3963416)

NVIDIA announced a partnership with Boston Dynamics to push these robot models into factories (0:59:19, l3963417)

Reports described workers paid to wear camera headbands while cleaning, stitching, and folding to train humanoid-robot dexterity models (1:07:49, l3963420)

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
Amit Shivatsa / Amish Yodsagoti	Same recurring Outskill host across episodes; name transcribed inconsistently by ASR (likely "Amit Shrivastava" or similar)
Veymour	Waymo, the self-driving robotaxi company
Norm Shazar	Noam Shazeer, AI researcher (Character.AI co-founder, later Google, reported to be joining OpenAI)
z dot a i / zed dot a i	z.ai, maker of the GLM 5.2 open-weight model
11 laps	ElevenLabs, the voice-AI company
MotionBricks / Motionpix / MotionVox	A motion-generation AI tool for 3D animation demoed in the Jun 24 episode; exact product name unclear/as-heard

True on recording day — verify before relying

T
h
e
s
e
f
o
u
r
e
p
i
s
o
d
e
s
a
r
e
i
n
f
o
r
m
a
l
,
l
i
v
e
-
s
t
r
e
a
m
e
d
n
e
w
s
r
o
u
n
d
u
p
s
r
e
c
o
r
d
e
d
J
u
n
3
-
2
4
,
2
0
2
6
,
d
e
l
i
v
e
r
e
d
f
r
o
m
t
h
e
h
o
s
t
s
'
p
e
r
s
o
n
a
l
i
m
p
r
e
s
s
i
o
n
s
o
f
h
e
a
d
l
i
n
e
s
,
b
e
n
c
h
m
a
r
k
s
,
a
n
d
s
c
r
e
e
n
d
e
m
o
s
r
a
t
h
e
r
t
h
a
n
v
e
t
t
e
d
r
e
p
o
r
t
i
n
g
.
C
o
m
p
a
n
y
n
a
m
e
s
,
m
o
d
e
l
n
a
m
e
s
(
e
.
g
.
,
"
F
a
b
l
e
5
"
v
s
.
"
M
y
t
h
o
s
"
)
,
p
r
i
c
i
n
g
,
v
a
l
u
a
t
i
o
n
s
,
a
n
d
b
e
n
c
h
m
a
r
k
c
l
a
i
m
s
a
r
e
a
s
-
h
e
a
r
d
a
n
d
w
e
r
e
n
o
t
i
n
d
e
p
e
n
d
e
n
t
l
y
v
e
r
i
f
i
e
d
;
s
e
v
e
r
a
l
i
t
e
m
s
(
t
h
e
F
a
b
l
e
5
/
M
y
t
h
o
s
b
a
n
,
S
p
a
c
e
X
'
s
I
P
O
/
v
a
l
u
a
t
i
o
n
,
t
h
e
C
u
r
s
o
r
a
c
q
u
i
s
i
t
i
o
n
,
A
n
t
h
r
o
p
i
c
'
s
I
D
-
v
e
r
i
f
i
c
a
t
i
o
n
p
o
l
i
c
y
)
w
e
r
e
e
x
p
l
i
c
i
t
l
y
d
e
s
c
r
i
b
e
d
b
y
h
o
s
t
s
a
s
s
t
i
l
l
d
e
v
e
l
o
p
i
n
g
a
t
r
e
c
o
r
d
i
n
g
t
i
m
e
a
n
d
m
a
y
h
a
v
e
s
i
n
c
e
c
h
a
n
g
e
d
,
b
e
e
n
r
e
v
e
r
s
e
d
,
o
r
b
e
e
n
c
l
a
r
i
f
i
e
d
.
T
r
e
a
t
a
l
l
f
i
g
u
r
e
s
,
d
a
t
e
s
,
a
n
d
p
r
o
d
u
c
t
c
l
a
i
m
s
i
n
t
h
i
s
d
i
g
e
s
t
a
s
a
s
n
a
p
s
h
o
t
o
f
w
h
a
t
w
a
s
s
a
i
d
o
n
a
i
r
,
n
o
t
c
o
n
f
i
r
m
e
d
f
a
c
t
.