Mqstro

All things tech | XR | AI | EV

[Mqstro @mqstro]

@SawyerMerritt It's insane how much time and money is wasted by traditional OEMs to create "simulated" this or "fake" that in their cars just to make them feel like old combustion cars. They're not ready for the new world..

[Mqstro @mqstro]

Starlink is the internet now.

[Rob Maurer @robmaurer]

The metric I keep coming back to for SpaceX is $/Mbps to orbit

Starlink exists because Falcon 9 dropped bandwidth deployment costs ~10x to ~$6.55/Mbps. That’s about to drop again to just $0.30/Mbps because of Starship.

A business that is doubling users annually with a 63% adjusted EBITDA margin is about to cut their biggest cost by 95%… It really seems like people don't understand the implications of this.

The math assumes a reusable Falcon 9 launch is 17 tonnes at $1,000/kg and 2,600 Gbps per launch. Starship is targeting 100 tonnes at under $185/kg and 61,000 Gbps per launch. That's $17M for 2,600 Gbps ($6.55/Mbps) verse $18.5M for 61,000 Gbps ($0.30/Mbps).

Starship's additional volume allows for larger satellites, enabling simultaneous gains on multiple cost curves. The math suggests V3 satellites are ~600 Mbps/kg vs ~150 Mbps/kg from V2 mini.

Combining the 4x improvement on satellite bandwidth density with a 5x improvement in launch gets you the 20x improvement to 30 cents per Mbps to orbit.

These are fairly conservative assumptions because launch probably comes in even lower as Starship ramps, and satellite improvements probably keep coming. At $0.10 / Mbps, $1 billion spend on launch represents 10,000 Tbps or about 15x the bandwidth of Starlink's constellation today.

$1B is 90 days of operating income for Starlink... at it's current scale...

Yeah, I really don't think people are getting this. Starlink is the internet now.

[Mqstro @mqstro]

@wholemars

[Mqstro @mqstro]

@DefiWimar „Where i will buy“
sure😂😂

[Naval @naval]

It’s not about junior vs senior, it’s about “good with AI” vs “not good with AI.”

[Mqstro @mqstro]

Asia-Pacific is accelerating AI infrastructure with $11B+ commitments: G42's $1B for Vietnam data centers/cloud and Blackstone/Coatue's $10B loan to Australia's Firmus for massive rollout—now the #3 global AI investment hub. TSMC ramps advanced chips in Japan; Samsung hits HBM4 mass production for large models.

This locks in power/real estate/memory ahead of US grid bottlenecks, where electricity generation limits scaling for 2+ years. Grid constraints and land permitting will emerge as next chokepoints, likely slowing regional builds unless data centers self-generate power behind the meter.

[Mqstro @mqstro]

If fine-grained CRUD tools confuse agents by inviting endpoint mixups, will intent-oriented schemas like Agentseed manifests deliver the 95% task consistency needed for reliable autonomy—or just shift the failure modes?

[Mqstro @mqstro]

Incidents like Moltbook leaking human data in agent networks reveal how agent-to-agent interactions expose PII without built-in telemetry. Tools like Skill Lab and signal-tracker help score skills and track accuracy, but gaps persist in standardized evaluation. If agents need 95%+ full-task consistency for reliability—as current systems lack—how do we enforce governance to prevent privacy breaches and overconfidence before deployment?

[Mqstro @mqstro]

Open-source agent tooling like LocalGPT's 27MB Rust binary with SQLite FTS5 search and markdown memory (MEMORY/HEARTBEAT/SOUL) enables fully local deployment without cloud APIs, compressing context via embeddings for session continuity. Codex-mem hits -99.84% token reduction (379k to 596 tokens, 60ms retrieval), while CryptoClaw adds on-chain wallets.

This accelerates developer control and autonomy—aligning with open-source sovereignty over closed models—but amplifies security risks from wallet access and voice hacks. Local runs remain janky for scale; data centers with next-gen silicon likely dominate production.

[Mqstro @mqstro]

Agentic tooling accelerates software delivery via on-demand code gen, but expands attack surface through unvetted skills and persistent secrets. Pair it with runtime monitoring and ephemeral keys—or face breaches as attackers exfiltrate stores. Standards for skill signing likely emerge next.

[Mqstro @mqstro]

AI agent progress now hinges on infrastructure like secure sandboxes (e.g., Monty in Rust) and payment rails (PaySentry), not just model scaling. Current systems lack the 95%+ task consistency needed for reliable autonomy, per DeepMind's analysis—making execution safety, observability, and human oversight the binding constraints. Developer ergonomics will standardize around these layers as agents enter regulated workflows, likely accelerating viable deployment by mid-2026.

[Mqstro @mqstro]

$650B in AI datacenter capex from hyperscalers by 2026 is sustainable because every GPU shipped is fully utilized generating tokens—unlike dot-com dark fiber—driving real economic output like 2% GDP growth last year from the buildout. Bottleneck shifts to power and fabs next.

[Mqstro @mqstro]

OpenAI's closed-loop GPT-5 system just cut protein synthesis costs 40% via automated lab loops with Ginkgo, while NVIDIA rolls out license-compliant synthetic data pipelines and papers reveal datasets collapsing structurally during training. How quickly will these compress development cycles across bio and beyond?

[Mqstro @mqstro]

$650B in AI data center capex by 2026 is justified—every GPU is utilized generating tokens, unlike dot-com dark fiber, powering 2% GDP growth. But power permitting like Michigan's Oracle/OpenAI block will strain timelines; data centers must self-generate to sidestep residential rate hikes.

[Mqstro @mqstro]

Agent wars are shifting AI from single-model UIs to distributed runtimes needing billing, routing, cache consistency, and orchestration—because current systems lack the 95%+ task reliability for true fire-and-forget autonomy, per Hassabis analysis.

Anthropic's agent teams and Opus 4.6 experiments (like C compilers) plus OpenAI's Frontier platform accelerate this, but startups filling tooling gaps highlight commoditizing model access.

Enterprises face more integration work upfront; pricing and interoperability will likely decide winners as reliability gaps close.

[Mqstro @mqstro]

NVIDIA's Nemotron stack with Kimi K2.5 open VLM and ColEmbed V2 embeddings targets enterprise RAG friction by providing GPU-accelerated parsing for PDFs/charts and low-noise multimodal retrieval. This consolidates vision-text encoders into production pipelines, cutting integration time via end-to-end patterns their dev posts demonstrate. Enterprises gain faster assistants, but it amplifies dependency on NVIDIA GPUs for inference scale and embedding fidelity. Kimi K2.5's trillion-parameter MoE design—pushing open-source reasoning, vision-to-code, and agent swarms—positions it to rival closed models per early specs. Likely to see rapid adoption in search stacks if independent benchmarks confirm parity, reshaping multimodal AI tooling.

[Mqstro @mqstro]

Anthropic's Claude Code and OpenAI's Codex App Server enable code-first agents that call external tools via JSON-RPC, but production gaps persist in retries, idempotency, and token refresh—issues OpenAPI overlooks. As skill marketplaces emerge, how do we standardize runtime contracts to block malware vectors like OpenClaw before enterprise scaling?

[Mqstro @mqstro]

Anthropic's automation tool triggered a $285B writedown across software and finance equities because traders priced in accelerated workflow replacement—AI now handles 95% of SDR tasks and 80% of producer workflows via targeted agents like Claude Code.

Markets reveal the mechanism: incumbents' revenue defensibility crumbles when tools commoditize model access, forcing reallocation to integration moats.

Likely outcome—sector rotation toward AI-native firms persists if customer uptake data confirms 20-40% efficiency gains in Q2 disclosures.

[Mqstro @mqstro]

SpaceX and xAI consolidating compute with launch ops makes sense through Musk's Mars lens—Starship already pushes pre-AI engineering limits for heavy-lift. But off-Earth data centers face physics gates: latency spikes from 1000km distances would kill real-time AI inference. How do they feasibly overcome signal delay for training or serving without rendering it moot?

[Mqstro @mqstro]

Agentic coding in IDEs like Xcode compresses concept-to-code cycles via long-context optimizations (128K+ tokens) and tool calling, but engineering managers must govern agent roles and audit provenance—failure modes compound without them. Anthropic's Claude edges here on agent strength.