Meng Li

AI Analysis and Stock Analysis, I share AI insights at

[Meng Li @dlimeng1]

Claude Code fast mode speeds up Opus up to 2.5x for higher token prices, best enabled at the start of a session and only when you're actively waiting on it. Do you usually leave it on all day?
#Claude #ClaudeCode #Opus
https://t.co/7Ka74LPstR

[Meng Li @dlimeng1]

Meta’s new Muse Wonder models close the gap with top AI rivals, but coding and 3D demo results are mixed. Can Meta turn expensive talent into a profitable personal agent?
$META #Muse #Wonder
https://t.co/E8sJhxAwYv

[Meng Li @dlimeng1]

AMD launches Helion rack reference design for agent workloads, built on OCP specs, with flexible CPU, GPU, power and cooling per task. Two options: CPU only racks with 6th Gen EPYC, up to 32768 cores for large scale agent execution linked to separate inference infra. Integrated racks combine EPYC nodes, Instinct MI350P GPU nodes and general compute for enterprises. A token router splits calls between frontier and open models. Air and liquid cooling supported. AMD plans to contribute Helion specs to OCP, is working with partners on rack configs, launch expected November 2026.

What does this mean for Nvidia in the agent infrastructure race?

$AMD

https://t.co/3LdlyAzOma

[Meng Li @dlimeng1]

Microsoft unveils Maia 300, its third-gen in-house AI inference system
Production deployment planned for 2027, scaling Azure Maia to multi-gigawatt capacity
Designed for long-context inference, low-latency chat and multi-step agent tasks
Microsoft expects 40% more tokens per megawatt than mainstream GPU systems
Rack-level token throughput up to 7x Maia 200 in full 96-accelerator config
Multi-chiplet architecture with compute chiplets on TSMC N2P process
8 HBM4 stacks per accelerator, 18TB/s peak memory bandwidth, 950W TDP
96 accelerators per rack at 167kW, 27.6TB total HBM capacity
Maia 200 already serves Copilot and other production workloads in US, Europe, Australia
GitHub Copilot and Microsoft Foundry to onboard by end of year

Will Maia 300 make Microsoft less dependent on Nvidia for AI inference?
$MSFT
https://t.co/GuUkWs17qh

[Meng Li @dlimeng1]

DeepSeek is reportedly leading a seed round of tens of millions of USD in robotics AI startup Winverse AI. The deal would value Winverse AI at about 400 million USD post money. Meituan's Long-Z Investments and others have committed. Winverse AI was founded in 2026 by Hu Han and is still in stealth. Hu was previously chief scientist at Tencent's vision language and embodied foundation model team, and worked at Microsoft Research Asia and Baidu. Its AI targets robots, not general chat apps.

This may be one of DeepSeek's larger startup investments since its 2023 founding. The deal is not final and terms may change. Hu Han, DeepSeek and Long-Z did not respond to Bloomberg requests for comment.

Will DeepSeek move further into robotics AI investment?

#DeepSeek

https://t.co/1G7J8Cwn8a

[Meng Li @dlimeng1]

Broadcom CEO Hock Tan said at a Singapore panel on Oct 12 that capital will become the biggest constraint for companies as AI continues to develop. He said Broadcom is overcoming wafer and memory shortages, and the next major bottleneck will shift to capital. On AI commercialization, he said Anthropic and OpenAI have proven that deploying the latest models as AI personal assistants can generate substantial revenue.

What does this shift toward capital as the main constraint mean for smaller AI companies?
$AVGO #OpenAI #Anthropic

https://t.co/gj2yZengjb

[Meng Li @dlimeng1]

Keep the Mermaid source, not just the flowchart image. What do you save when AI builds your diagram?
#Mermaid #flowchart #image #diagram #AI
https://t.co/GPFMD6rqlK

[Meng Li @dlimeng1]

ArtCraft is an open-source AI creation workbench that lets you build 3D scenes before generating, making AI images controllable. Could this scene-first approach become the future of AI creation?
#ArtCraft #AI #3D #scenes #images
https://t.co/ZJ79XYlFHD

[Meng Li @dlimeng1]

Is NextPower (NXT) a viable investment, or does the tug-of-war between its fundamental value and technical setup signal risk?

This analysis evaluates the business model of NextPower to determine its position in the current market. We move beyond the surface-level narrative to examine the company's financials and intrinsic value, weighing them against the realities of its current price action. By contrasting the firm's operational strength with the chart potential, we aim to clarify whether the stock warrants a place in your portfolio.

Beyond the operational review, we evaluate the firm through a detailed financial model. By analyzing the Elliott Wave setup alongside current balance sheet trends, we help you decide if the risk profile remains unfavorable at these levels. We specifically review the balance scale between value and price to determine if the stock is correctly positioned.

Subscribe for weekly stock market analysis and detailed financial breakdowns. Let us know in the comments if you are following this company or waiting for a better entry point.

#ai #datacenter #nuclear #energy #solarenergy $NXT $TE $SMR

https://t.co/11sExU3rMG

[Meng Li @dlimeng1]

Nextpower at $85 trades above its $55-$75 intrinsic value range, with strong backlog and net cash but no margin of safety. Is it worth buying now?
$NXT $TE $SMR
https://t.co/baq9nG7dhj

[Meng Li @dlimeng1]

AI is accelerating drug discovery, but the real bottleneck and investment opportunity may be upstream lab tools and data companies. Which of the four companies do you think has the strongest long-term potential?
$RXRX $RVTY $TWST
https://t.co/LlCVzbosgW

[Meng Li @dlimeng1]

InP laser capacity can't meet demand as CPO and pluggable modules grow together, with tight supply likely lasting until 2028. Can the supply chain expand fast enough?
$LITE $COHR $AAOI
https://t.co/StzhdRkOyd

[Meng Li @dlimeng1]

Claude Code mobile lives in the Claude app's Code tab, letting you check progress and direct local or cloud sessions from your phone. Have you tried /mobile yet?
#Claude #ClaudeCode #mobile
https://t.co/S9waDdhVrg

[Meng Li @dlimeng1]

OpenClaw v2026.10.1 adds Skills search and MCP Apps but user fatigue over constant updates grows. Where is OpenClaw heading now?
#OpenClaw #Skills #MCP #Apps
https://t.co/uYz210rG5L

[Meng Li @dlimeng1]

Karpathy: only ~5,000 people can access frontier internal AI. ~6B have barely touched LLMs. ~1B to 2B use free ChatGPT. ~20M use frontier LLMs in math and coding. In labs, thousands of agents run for weeks, build software, find zero days, and push science.

A bit unexpected?

#AI #ChatGPT

https://t.co/1G7J8Cwn8a

[Meng Li @dlimeng1]

Google now lets candidates use Gemini in interviews.

New round: read, debug and judge AI written code.

AI fluency is the new metric.

Canva and Meta are testing the same.

Writing code is easier. Reading it is not.

Would you hire someone great at prompting but weak at writing code?

#Google #Gemini

https://t.co/GeUBPBgz4F

[Meng Li @dlimeng1]

Zhipu GLM 5.3 Highspeed is rolling out in Zcode 3.15.1 for random subscribers, per @leo114119.

Speed hits 490 tokens/s. GLM 5.3 FlashX topped 200.

Pi lists 1M context and 128k output. Reddit ties it to Omen Alpha.

Still random. I could not test it yet. What are you seeing?

#Zhipu #GLM

https://t.co/t7ErJKevny

[Meng Li @dlimeng1]

AI is now sweeping through mathematical proofs. A math enthusiast used GPT-6 Astra to push a 57 year old problem forward. The question: given N unit disks, how do you pack them into the smallest possible larger circle? Mathematicians found a good arrangement in 1969 but could not rigorously prove it optimal for 57 years. nor used GPT-6 Astra to rigorously prove the remaining cases N=15, 16, 17, 18, and 20. Previously only N<=14 and N=19 were verified. N=20 is not the AI limit, just where nor stopped. The proof approach: push outer disks to the boundary, use angle constraints to eliminate arrangements, let a computer enumerate remaining cases, use geometric rigidity to upgrade near optimal to exact, handle N=15 with an odd cycle argument, then restore the original arrangement. nor says his human contributions included feasibility, guidance, fighting a bad web UI, fixing details, manual verification, and rechecking with Astra and Opus 5.5. Others are using AI on the same problem. Could this mark a new era where AI makes advanced math more accessible and collaborative?

#GPT #Opus

https://t.co/gj2yZengjb

[Meng Li @dlimeng1]

Google's Gemini 4 Argon beats Claude Opus 5.5 on some coding tests but trails GPT-6 Astra elsewhere, and it is worth testing for long-form writing. Will it actually hold up in real workflows?
#Google #Gemini4 #Claude #Opus #GPT6 #coding
https://t.co/w9Y47QDImN