Vivek Varma Siruvuri

Architect | AI Builder | 20+ yrs turning complex tech into real outcomes | Writing about AI, leadership, and the future of how we build |

[Vivek Varma Siruvuri @svivekvarma]

Claude Opus 4.7 dropped April 16.

Everyone posted the wins. Here's the full picture — what progressed, what regressed, and what breaks your code. 🧵

[Vivek Varma Siruvuri @svivekvarma]

🟡 STILL A BILL

NO FAKES Act: federal right over your voice + likeness, 70 years. White House endorsed it by name. SAG-AFTRA, Google, OpenAI, Disney all backing it.

Still in Senate Judiciary Committee.

The EO is real. The law is real. The Framework is a wish list. The NO FAKES Act is a bet. Know which is which.

[Vivek Varma Siruvuri @svivekvarma]

Everyone's talking about "US AI regulation." Most can't name what's actually law vs. a memo.

Here's the full stack, right now 🧵

[Vivek Varma Siruvuri @svivekvarma]

🔵 POLICY MEMO (not law)

The March 2026 White House Framework is a recommendation to Congress — not a regulation.

Key position: no new AI agency. Route everything through FTC, SEC, FDA, DOE. Congress hasn't acted.

[Vivek Varma Siruvuri @svivekvarma]

Vision and coding gains are real.

Long-context retrieval collapse is also real.

It depends on your stack.

[Vivek Varma Siruvuri @svivekvarma]

Claude Opus 4.7 dropped April 16.

Everyone posted the wins. Here's the full picture — what progressed, what regressed, and what breaks your code. 🧵

[Vivek Varma Siruvuri @svivekvarma]

⚡ What breaks your code:

→ temperature / top_p / top_k — removed entirely, 400 if you pass them
→ Extended thinking — only "adaptive" mode supported now, "enabled" + budget_tokens = 400

Check your API calls before you deploy.

[Vivek Varma Siruvuri @svivekvarma]

* Cloudflare Workers AI = open-source models (Llama). Dollar figures aren’t capability-comparable to Claude or GPT-4o-mini.

But: Cloudflare AI Gateway can proxy Anthropic/OpenAI calls and cache those too. They’re not alternatives — they stack.

Full repo with all 4 demos:
https://t.co/pAPAl3dBsh

[Vivek Varma Siruvuri @svivekvarma]

Same workload. 4 providers. Here’s what AI caching actually saved at 1M requests.

The results — and one surprising finding 🧵

[Vivek Varma Siruvuri @svivekvarma]

⏱️ The surprising finding — hours freed at 1M requests:

Cloudflare: 432 hrs
Bedrock: 379 hrs
Anthropic: 317 hrs
OpenAI: 218 hrs

Cloudflare wins on time saved — not because of the model, but because gateway hits return in 498ms vs 2,100–2,500ms for the others.

[Vivek Varma Siruvuri @svivekvarma]

Stack semantic → prefix → KV. They're a cascade, not alternatives.

Next: same stack, 4 providers, real API calls, real costs.
At 1M requests — one provider saved 432 hours of compute time.
It's not who you'd expect.

Numbers and stats dropping soon.

[Vivek Varma Siruvuri @svivekvarma]

Use Generative AI caching, unless you want to throw away some $$$$$

3 layers. Most teams use zero. 🧵

[Vivek Varma Siruvuri @svivekvarma]

🗄️ Semantic Cache — intercepts before any API call is made

"Recommend a thriller" ≈ "suggest a suspenseful film"
Cosine similarity: 0.86 → return cached response

Cost: $0.00
Latency: <1ms vs 2,845ms baseline
No model called. No tokens spent.