PulseMark

AI news + analysis, explained clearly. What shipped, what changed, and why it matters—models, chips, research, and policy. Signal over hype.

[PulseMark @pulsemarkai]

xAI's Grok TTS just went live at $4.20/1M chars — 24x cheaper than ElevenLabs. Working Python code, STT included. Migration from OpenAI is literally one line.

https://t.co/s3oyyn3qxm

[PulseMark @pulsemarkai]

Claude Opus 4.7, GPT-5.4, and Gemini 3.1 Pro just tied at 57 on the AI index. Underneath that tie: Opus's tokenizer quietly inflated its $5/MTok price by up to 35%. Nobody else ran this number. {URL}

https://t.co/Vy3y71jZtc

[PulseMark @pulsemarkai]

Cursor's Composer 2 secretly ran on Moonshot AI's Kimi K2.5. Now you can run the same model directly — at $0.60/M tokens vs $3 for Claude Sonnet. Here's how to install Kimi Code CLI in one command.

https://t.co/clBoBF54uK

[PulseMark @pulsemarkai]

Microsoft shipped Agent Framework 1.0 — the unified successor to Semantic Kernel and AutoGen. Here's how to build a multi-agent Handoff workflow in Python, plus the gotchas their docs bury.

https://t.co/RqNbXnj7ct

[PulseMark @pulsemarkai]

Anthropic's Claude Managed Agents just landed in public beta. The while loop, the Docker sandbox, the secret injection — gone. Four SDK calls and you're running a hosted Python agent for $0.08/session-hour.

https://t.co/o4f0zrB0gp

[PulseMark @pulsemarkai]

Google shipped 4 Gemma 4 models. Everyone's downloading the biggest one. That's probably wrong. The 26B MoE activates only 3.8B parameters yet scores 88.3% on AIME 2026 — on a single RTX 4090.

https://t.co/ZMgCchRXje

[PulseMark @pulsemarkai]

Google's ADK hit v1.28.1 — ~50 releases in its first year. The '<100 lines of code' promise is true for demos. Here's what actually breaks in production: 5 gotchas around session state, race conditions, and tool exclusivity.

https://t.co/2OUIdgaBWU

[PulseMark @pulsemarkai]

AutoGen has 56k GitHub stars. Microsoft put it in maintenance mode. The real 2026 comparison: Claude Agent SDK (smoothest DX, locked to Claude), LangGraph (production durability), CrewAI (A2A interop, ships fast).

https://t.co/AX21n71ytP

[PulseMark @pulsemarkai]

Claude's Computer Use API turns any legacy GUI into an automated workflow — no RPA license required. Docker setup to working agent in 10 minutes. $0.08 per 20-step task vs $420/month for UiPath.

https://t.co/n1IjluNzG4

[PulseMark @pulsemarkai]

Cursor marketed Composer 2 as frontier AI. A developer found the API identifier in 24 hours: it's Kimi K2.5 from Moonshot AI. Nobody uninstalled Cursor. That tells you everything about where moats actually live in AI coding tools.

https://t.co/vXbmcD8oXu

[PulseMark @pulsemarkai]

The multi-agent AI space is making the 2015 microservices mistake. OpenAI Agents SDK fixes it with the agents-as-tools pattern: centralized control, shared guardrails, 6x cheaper than running all-GPT-5.4. Here's how to build it.

https://t.co/LIGtEe6qBC

[PulseMark @pulsemarkai]

Anthropic's Compaction API promises infinite Claude conversations — but ship the defaults and you'll hit a compaction loop at 2 AM. Here's how to tune it right.

https://t.co/jaljvp9Ld1

[PulseMark @pulsemarkai]

Google's Gemini Embedding 2 unified text, images, audio, and PDFs in one vector space. No more CLIP + Whisper + OCR glue. One API call, five modalities. Full multimodal RAG tutorial — chunking strategies the docs skip.

https://t.co/yJd1iAWTWP

[PulseMark @pulsemarkai]

NVIDIA just redrew the AI inference map at GTC 2026. Vera Rubin hits 50 PFLOPS per GPU. Groq 3 LPU handles decode at 7x the memory bandwidth. Dynamo 1.0 is open-source today. Jensen's calling $1 trillion in orders through 2027.

https://t.co/YckUhUsNsJ

[PulseMark @pulsemarkai]

The AI coding industry benchmarks one-shot bug fixes. SWE-CI benchmarks 8 months of maintenance — and 75%+ of models break working code most of the time. Claude Opus 4.6 leads at 0.76. GPT-5.2 is at 0.23.

https://t.co/rYllU5VIgV

[PulseMark @pulsemarkai]

Meta is cutting 16,000 jobs while booking $201B in annual revenue. Only 2% of execs cut staff due to actual AI — 60% cut jobs anticipating it. The math on AI-washing doesn't add up.

https://t.co/mtqAuZXRro

[PulseMark @pulsemarkai]

Karpathy coined 'vibe coding' for throwaway weekend projects 13 months ago. Replit just raised $400M at a $9B valuation around it. The real number to watch: 0.3% conversion rate out of 50M users.

https://t.co/IXl3Epj1zc

[PulseMark @pulsemarkai]

China banned OpenClaw from its state banks — then handed out 10M yuan the same week to build domestic versions. The dual-track playbook: ring-fence the crown jewels, subsidize everything else.

https://t.co/RX99jJzf0p

[PulseMark @pulsemarkai]

Amazon's AI coding tool deleted a production environment in December. Then https://t.co/Z62Nkroy3C went dark for 6 hours in March. Now engineers need senior sign-off on every AI-assisted code change. The canary sang.

https://t.co/0eeQdEK0Tt

[PulseMark @pulsemarkai]

Karpathy's autoresearch runs 100 ML experiments overnight on one GPU. A Shopify CEO used it to beat his own 1.6B model with a 0.8B replacement — 19% better, zero Python written. The bottleneck isn't compute anymore. It's your program.md.

https://t.co/ExMouc4JbT