Sharapov Andrei

Building CryPrice — DeFi risk intelligence.

Wallet risk • Health Factor • Liquidations • Market events

[Sharapov Andrei @AdrasSad]

I build backend infrastructure for crypto products.

Node.js, PostgreSQL, Redis, Telegram bots, wallet monitoring, DeFi alerts, Aave risk tools, VPS deployments.

Currently open to short-term paid tasks.

If you need reliable crypto/backend infra — DM me.

[Sharapov Andrei @AdrasSad]

Not 1:1 in Cursor.

Fast/Expert ≈ Grok Effort low vs high/xhigh
Auto ≈ Cursor Router
Heavy ≈ parallel agents, not a model mode
Build ≈ Agent, or Plan → Build

Cursor “Fast” is a 2× speed tier. Max Mode is context, not Expert.

Is Heavy coming to Cursor as a first-class toggle, or will it stay Build-only?

[Sharapov Andrei @AdrasSad]

Never saw the point of running OpenAI models as the default path inside Cursor.
Cursor is not a model. It is a harness: repo context, agent loop, diffs, rules, project memory. Session quality is almost always about how a model behaves inside that loop - tool-calling stability, instruction following, long-context discipline, cost per iteration. On that profile GPT in Cursor has not been the best default for a long time. Claude usually holds the agent loop better. Composer / Cursor’s router is cheaper and fitted to the product. After the SpaceX deal, Grok is the first-party path. OpenAI is also competing with Codex, so it was supplying a model into someone else’s IDE while building a replacement for that IDE.
That is why the Nov 12 cutoff sounds loud and is operationally small. If GPT was ~5% of Cursor traffic, this is not a platform loss. It is OpenAI cleaning a competitor’s channel after a change-of-control clause. Tab, Cloud Agents, Auto routing, CLI and Cursor’s own API were never going to run on your OpenAI key anyway.
GPT inside Cursor only makes sense in two cases: a narrow workload where that specific OpenAI line is actually better on your codebase, or BYOK when you already pay OpenAI separately and want one key in local Chat/Agent. As the default model in Cursor, it was a weak choice long before this announcement.

[Sharapov Andrei @AdrasSad]

@avax @Delphi_Digital $1.2B live or $2.7B migrated? And can any of those Progmat tokens move to a wallet that isn’t on the Japanese whitelist?

[Sharapov Andrei @AdrasSad]

Zondacrypto: The Real Risk Was Never Crypto — It Was Custody https://t.co/garNnvbIha

[Sharapov Andrei @AdrasSad]

Your Ledger can show you one transaction and sign another.

No funds were reported stolen, and the bug is patched. But the bigger lesson is worse: hardware wallets aren’t the final security layer anymore.

Clear Signing is.

Would you trust your signer without it?

[Sharapov Andrei @AdrasSad]

Ledger wasn’t hacked. The trusted display was.
OneKey reproduced an already-patched Ethereum app bug: the screen can show one tx and sign another. No keys extracted. No confirmed thefts. Update to 1.22.3+ and verify the version on the device.
If you haven’t checked the device screen, you’re underwriting a lab exploit.
Have you verified the app version on-device?

#CRYPTONewsTalk

[Sharapov Andrei @AdrasSad]

@Im_IrushiK Sol only for planning and review. Luna for the actual work. Otherwise the 5 hours burn while the agent is still deciding which folder to start in.

[Sharapov Andrei @AdrasSad]

@avax Mr. Wonderful looking at Avalanche is already a statement. Curious what he actually thinks of the network.

[Sharapov Andrei @AdrasSad]

@cb_doge Reset is nice. The real unlock is that Grok Bot can draft the work, but still can’t ship the post for you. That’s the correct default.

[Sharapov Andrei @AdrasSad]

Smart framing: model + harness, and only the harness can be audited. That’s the real product. Coding-agent comparisons miss it. Security-coworker comparisons should be harsher - these are workflows, not a scanner. I’ll take OpenWorker for sensitive repos and shift-left reviews; I won’t leave Cursor/Claude Code for it.

[Sharapov Andrei @AdrasSad]

Hugging Face just proved the open-weight paradox: closed models’ guardrails locked defenders out while an autonomous OpenAI agent ran 17k+ actions against their infra.
DeFi teams using AI agents or HF models - your threat model just changed. Self-hosted fallback is no longer optional.
What are you running for incident response when the model says “no”?

[Sharapov Andrei @AdrasSad]

This reads like slick PR hype mixing real trends with tech buzzwords. The concepts are solid (e2e, low latency, conversational RL), but 'parallel reasoning while speaking' directly contradicts autoregressive mechanics, and stats like 0.7s TTFA or '0.4x tokens' remain unverified.

Until there are open benchmarks and independent API tests, it's just great marketing.

Have you actually tested the API on live calls yet, or are you just going off the pitch deck?

[Sharapov Andrei @AdrasSad]

Most agent token waste and wrong tool calls come from the same place:
You gave the model 40 tools when it needed 6.
Optimize in this order:
1. Ruthlessly cut the toolset by role
2. Deferred loading by default
3. Write sharp, distinctive tool descriptions
4. Add tool search only if you still have >15 tools
Everything else is secondary.

[Sharapov Andrei @AdrasSad]

Do you believe that transparency alone (documenting the caveats) is enough to change the culture of how benchmarks are used in the field, or are you planning more active mechanisms - for example, a public ‘reliability’ rating for benchmarks, penalties in leaderboards for known leakage, or incentives to create truly leakage-controlled alternatives?

[Sharapov Andrei @AdrasSad]

Intel and AMD use classic SMT: two threads share the same core resources. This improves utilization but creates contention and latency noise.
NVIDIA took a different path with Vera - Spatial Multithreading. Core resources are physically partitioned. You can give the entire core to one thread or split it into two isolated ones.
For Grok’s agent loops (tool calling + code execution) the difference is critical:
• Intel HT - ~62% efficiency
• AMD SMT - ~68%
• NVIDIA Spatial - ~91%
That’s why Vera is built for agents, not just for “more threads.”

[Sharapov Andrei @AdrasSad]

@teslaownersSV Grok 4.5 again #1 on actual hard engineering tasks from real PRs. Not synthetic benchmarks - the stuff that actually ships. 🔥

[Sharapov Andrei @AdrasSad]

The more I use AI for development, the more I realize:

AI doesn’t replace engineers.

It replaces engineers who only know what to code.

The real advantage now is knowing:
- what to build
- why to build it
- how to structure it
- how to verify the result
- and when AI is confidently wrong.

Coding is becoming cheaper.

Engineering judgment is becoming more valuable.

[Sharapov Andrei @AdrasSad]

@AndrewYNg Evals are the only skill that actually separates systems from expensive theater.
Everything else is already table stakes.

[Sharapov Andrei @AdrasSad]

@StephenZeiner Fair point.
Permits alone won’t stop militant local opposition.
$KEEL needs a clear “Fact or Fiction” on water, noise, traffic and real community benefits, especially with DEP still open.

[Sharapov Andrei @AdrasSad]

More loyal / supportive version:
$99M stopped is solid progress.
Pre-payment screening against the Death Master File is exactly the kind of practical win DOGE was pushing for. Better to catch it before the money leaves than chase it later.
Glad Bessent and Treasury actually delivered on this. More of this, please.