Tian Zhang (Brice)

Daily Update AI and Robot Trend news and Stocks and my Insights

[Tian Zhang (Brice) @BriceTian]

$GOOGL Daily — Oct 26
$316.80 | -1.1%
LT 🟢74 MT 🟡56 ST 🔴52
Key: $316.60 support / $324.46 resistance
Fell while the Nasdaq rose 0.8%; closed under $317.48, the last marked support.
Posture: Hold
23.7x fwd, consensus PT $425 — but no support below until Wed's print + FOMC.

[Tian Zhang (Brice) @BriceTian]

$META Daily — Oct 26
$630.56 | -1.88%
LT 🟢64 MT 🟡52 ST 🔴58
Key: $627.80 support / $632.45 resistance
Up to $95B equity raise for AI capex; buybacks zero in 2026, Q3 FCF negative.
Posture: Hold
Ad engine intact, but a $60B ATM sells into my own recovery thesis.

[Tian Zhang (Brice) @BriceTian]

$AMD Daily — Oct 26
$531.22 | +2.8%
LT 🟢75 MT 🔴58 ST 🟡55
Key: $527 support / $542.61 resistance
Anthropic — AMD's 2GW MI450 customer — filed to IPO on $16.3B of 1H revenue.
Posture: Hedge
Sector beta, not AMD demand. 48x fwd into a hawkish Fed — hedged, not naked.

[Tian Zhang (Brice) @BriceTian]

$NVDA Daily — Oct 26
$216.03 | +2.40%
LT 🟢65 MT 🟡58 ST 🟡52
Key: $210 support / $218.9 resistance
SpaceX closed $40B of DEBT for NVDA GPUs — best demand news of the correction, and NVDA still lagged SMH.
Posture: Hold
17.9x fwd is fair; FOMC + MSFT/GOOGL on Oct 28 decide it.

[Tian Zhang (Brice) @BriceTian]

Into tomorrow: FOMC day one. A hold is priced, so the tape trades December odds. Tech leadership holds = range intact; tech fading with small caps = last week's downtrend resumes.

Megacap earnings (MSFT, GOOGL, AMZN) midweek.

⚠️ Information roundup, not investment advice.

[Tian Zhang (Brice) @BriceTian]

Levels: S&P 500 ~7,681 recovers last week's drop but sits roughly 1.5% below the 7,800 zone where it logged its first record close on Oct 5. That band still caps the range.

Nasdaq 100 reclaimed 29,300; Composite ~26,150. Russell 2000 ~2,785 remains the laggard.

[Tian Zhang (Brice) @BriceTian]

Unusual moves, with catalysts:

SGRY +46% — take-private at $23.50/sh
XNCR +22% — Phase 1 data
RRC +15% / EQT -3.6% — >$50B all-stock merger
TSLA +6% — Barclays to Overweight, $450
XP -48%, NU -20%, EWZ -22% — Lula wins Brazil runoff
AVBP -23%, FICO -11%

[Tian Zhang (Brice) @BriceTian]

Rotation: Tech (XLK) +1.6% led, on chips and optical networking. Industrials and consumer discretionary each +1%.

Comm services -1.2% was the lone drag (NFLX -3.7%); staples and health care slipped. 10 of 11 sectors green.

WTI $102.72 -0.5%. Gold $4,034 +0.7%.

[Tian Zhang (Brice) @BriceTian]

No US data today, so the bid was positioning into the Tue-Wed FOMC. CME pricing has ~91% odds of a hold Oct 28, with the real argument moved to December.

10y 4.98% (-2bp), 2y 4.71% (+2bp) — curve flatter. VIX -5.9% to 18.31. Treasury sold $69B in 2s at 1pm.

[Tian Zhang (Brice) @BriceTian]

[Before Market Close Update] — October 26

S&P 500 7,681 +0.8%
Nasdaq 100 29,306 +1.2%
Dow 52,349 +0.3%
Russell 2000 +0.7%

A chip-led tech rebound carried the tape into the FOMC, clawing back part of last week's slide. Levels are intraday, ~1pm ET.

[Tian Zhang (Brice) @BriceTian]

This is a one-line fix living in LeRobot's SmolVLA default and the π0/π0.5 guides. It's LIBERO-only though, single author, no real-robot runs. Open question: how much of the published VLA ablation record is measuring optimizer precision, not method?

[Tian Zhang (Brice) @BriceTian]

9 of 12 open VLA codebases fine-tune with bf16 weights and no fp32 master copy. A probe every 100 steps found ≥79% of AdamW updates silently rounded to zero. (Oct 22)

https://t.co/mUyHosdfdW

[Tian Zhang (Brice) @BriceTian]

The mechanism: bf16 keeps 8 significant bits, so any update below |w|/512 rounds back to the original weight. Adding an fp32 master copy on LIBERO-Object took π0.5 from 5.1% to 80.9% and SmolVLA from 4.3% to 49.3% at a 10x-lower LR.

[Tian Zhang (Brice) @BriceTian]

$TSLA Daily — Oct 26
$394 | +5.8%
LT 🟢73 MT 🟡58 ST 🟡55
Key: $386 support / $397 resistance
Barclays to Overweight, $450, on a SpaceX deal nobody has terms for.
Posture: Hold
First 200-day break in 14 sessions, but it's a merger premium, not earnings.

[Tian Zhang (Brice) @BriceTian]

It turns inference-compute control into a cheap head on a frozen model: 21.2x fewer params and 55x fewer GPU-hours than training price-specific checkpoints. Caveat: only 1.7B and 7B, one main baseline. Open question: does the answer settle this early at 70B?

[Tian Zhang (Brice) @BriceTian]

A linear probe on a frozen reasoning model's hidden states can tell when the answer is already settled and stop the generation: 41% fewer tokens on Qwen3-1.7B, 51% on OLMo-3-7B-Think, no accuracy loss. BELT (Oct 21)

https://t.co/qGkOjME6tA

[Tian Zhang (Brice) @BriceTian]

BELT splits the stopping decision in two. A probe reads frozen hidden states to predict the current and the eventual answer; a linear policy then weighs more thinking against its token cost. The targets ignore token price, so one probe refits to any budget.

[Tian Zhang (Brice) @BriceTian]

Setup: the Nasdaq sits more than 8% below its record, so leadership is the open question.

Above 7,812 on SPX, watch a retest of 7,844. Below 7,578, 7,483 comes into play. Falling oil and a 4.99% 10Y pull against each other.

⚠️ Information roundup, not investment advice.

[Tian Zhang (Brice) @BriceTian]

📈 Pre-Market Update — October 26

ES 7,672.75 +0.17% · NQ 29,141.25 +0.08% · YM 52,577 +0.35%

What matters most: this is Fed week. Futures are drifting up on cheaper oil rather than tech leadership, with Wednesday's FOMC sitting in front of Big Tech earnings.

[Tian Zhang (Brice) @BriceTian]

Levels.

SPX cash: resistance 7,782–7,812 (August high close and record close), then the 7,844 record. Support 7,578 (May high close), then 7,483 — the 23.6% retracement of the yearly range.

NDX: resistance 30,046 (August high close). Support 28,753, the September low.