Optivaize
281 posts
Optivaize
@Optivaize
We build practical AI systems. Insights, builds, and real-world AI.
انضم Şubat 2026
146 يتبع128 المتابعون
@IBuzovskyi Sales→ops handoff was the one we had to think hardest about. we ended up with structured payloads instead of letting the agent infer.
Would love to connect with you!
English
@Optivaize role-based MCP scoping is the right architecture.
founder sees everything, ops handles workflows,
team profiles stay in their lane.
the isolation between profiles keeps it clean.
no shared memory, no skill bleed across roles.
English
HERMES AGENT RAN AUTONOMOUSLY ON A $7 VPS
FOR 30 DAYS. 2,904 SESSIONS. 7,544 TOOL CALLS.
ITS OWNER STEPPED IN 18 TIMES. THAT'S IT.
2,879 cron jobs executed without input.
inbox management, background research,
automated workflows. all hands-off.
the setup:
the agent calls APIs, not the model itself.
no GPU needed.
minimum: 2 vCPU, 2GB RAM
(gateway + Telegram + cron)
recommended for heavy use: 4 vCPU, 8GB RAM
VPS options (real prices with IPv4 + tax):
→ Hetzner CX22: ~$7/month
2 vCPU, 4GB RAM, 40GB NVMe
→ Contabo VPS S: ~$7/month (annual)
4 vCPU, 8GB RAM, 50GB NVMe
→ DigitalOcean: ~$24/month
2 vCPU, 4GB RAM
→ IONOS: ~$2/month
cheapest entry, no annual lock
model provider (separate from VPS):
→ Nous Portal Plus: $20/month
$22 in credits, 300+ models,
web search, image gen, TTS, browser
one sub replaces all API keys
→ Grok via SuperGrok: $0 extra
OAuth, no API costs
→ bring your own keys:
Anthropic, OpenAI, OpenRouter
pay per token
the install:
1. spin up any Linux VPS
2. SSH in:
curl -fsSL https:/ /raw.githubusercontent.com/ NousResearch/hermes-agent/main/scripts/install.sh | bash
3. hermes setup (pick model + connect Telegram)
4. set your cron jobs
5. close the laptop
full Hermes setup guide in the article 👇
YanXbt@IBuzovskyi
English
@IBuzovskyi a handful, set up per role. founder profile has full access, ops and team profiles are scoped down depending on what they need. the role-based split matters more than the count for us, especially since each profile gets different mcp access.
English
@Optivaize exactly. the VPS is infrastructure, the value
is in what you point it at. teamleader + slack +
meeting notes is a solid stack.
how many profiles are you running across the company?
English
The US tried to block China from Nvidia chips. China responded by blocking Nvidia from China.
China made it official. AI chips are now on Beijing's "secure and reliable" list.
That means every Chinese government agency, state company, and strategic industry is legally required to buy domestic.
The approved stack:
→ Huawei Ascend 310 and 910
→ Alibaba T-Head Zhenwu M530 and M890
→ Biren, Hygon, MetaX, Moore Threads, Iluvatar CoreX
Nvidia is not on the list. AMD is not on the list. The Western AI stack is legally walled off from the second-largest economy on earth.
Lisuan also just shipped China's first fully domestic gaming GPU. It does not beat the RTX 4060 yet. That is not the point. The point is that it exists, sold out of 30,000 units, and got Microsoft WHQL certification.
What this means for businesses building on AI:
→ There is no longer one global AI stack. There are two.
→ If you sell in China or to Chinese partners, you need to know which chips and models they are required to use.
→ If you build on Western AI, your stack is going to evolve faster on cutting-edge models but stay locked out of half the world.
→ Western chip prices stay high because the biggest competing buyer just left the market.
Most companies treat AI like a tool. It is becoming infrastructure. And infrastructure decisions made now will be hard to reverse in 3 years.
What do you think will happen in the long run?
Will China overtake its competitors?
Is the top spot actually within reach, or will Nvidia stay unbeatable?
English
5.5 genuinely smashes specific stuff (long codebases, debugging, agent loops) and claude still owns design and planning. the discourse swing isn't really PR, new model launches always reset the conversation for a few weeks before the routing reality settles back in.
Sushil Kumar@Sushilk91
Why is everyone suddenly glazing GPT 5.5 and trashing Claude Code? I’ve been using both Opus 4.8 + GPT 5.5 (and even 4.7) daily. 5.5 is genuinely strong at a lot of things, but when it comes to deep planning and frontend/design work, it’s still nowhere close. Even with great skills, the outputs are often laughably bad. Is this organic or is OpenAI pushing the narrative hard?
English
running both daily, the split looks like what AWS literally tells customers in their own builder docs: claude for new builds, gpt for migrations and specific debugging workloads. the discourse always overcorrects on launches. one model winning everything was never going to be how this played out.
English
Why is everyone suddenly glazing GPT 5.5 and trashing Claude Code?
I’ve been using both Opus 4.8 + GPT 5.5 (and even 4.7) daily. 5.5 is genuinely strong at a lot of things, but when it comes to deep planning and frontend/design work, it’s still nowhere close. Even with great skills, the outputs are often laughably bad.
Is this organic or is OpenAI pushing the narrative hard?
English
@zekramu most serious AI work doesn't run on windows today, but that's why this announcement exists. enterprise IT is overwhelmingly windows-shop. nvidia and microsoft are betting the on-device agent shift happens in the place where most users actually are, not where the linux crowd is.
English
we've been running 80b+ models on a 128gb macbook for sovereignty clients all year. windows users finally get the equivalent option. the on-device agent push is going to define the next 18 months of AI hardware. the question is whether the software ecosystem catches up at the same speed.
English
Opus 5.0 - limit reached before you even open the app...
Prajwal@0xPrajwal_
Opus 4.5 - limit reached after 30mins Opus 4.6 - limit reached after 15mins Opus 4.7 - limit reached after 10mins Opus 4.8 - limit reached after 5mins Opus 4.9 - limit reached while prompting Continue this......
English
Hermes literally gets named in the announcement! Nous research saying RTX Spark + OpenShell gives hermes users a secure on-device environment.
We run hermes in prod and the local-on-device option keeps getting more credible.
Fall ship date though, current setups don't change today.
NVIDIA@nvidia
NVIDIA RTX Spark: a 1-petaflop superchip, the full CUDA and RTX ecosystem, and Windows-native agents. A new beginning for personal computers.
English
@bridgemindai yeah this is our setup in production too. our split: claude for building, gpt for parts where claude gets cautious or stuck debugging, deepseek for high-volume routine work lately. shifts per task. the savings start showing up once you stop forcing one model to do everything.
English
GPT 5.5 and Claude Opus 4.8 are best used together.
People are wasting time arguing over which one to pick.
I just finished a 5 hour vibe coding session running both in parallel.
Here's the strategy.
When Opus 4.8 got stuck on a bug it couldn't crack, I handed it to GPT 5.5 and it solved it.
When GPT 5.5 shipped a garbage UI, I had Opus 4.8 clean up the design.
Each model covers the other's weakness.
Opus 4.8 for design and intelligence.
GPT 5.5 for hard debugging.
Stop picking sides.
The builders winning right now run both.
English
@PaulGugAI from client work, the productivity story isn't keeping up with the valuation story yet. real gains, yes, but not at the level the equity weights imply. the capex side is way ahead of realized productivity. those reconcile eventually somehow, just not sure how yet.
English
S&P 500 new record highs (ATHs) tend to tell a different story each time. And here we are again.
Two points stick out for me:
- Only 20 of the 500 are at their own ATH. This is concerning in terms of overall health of the economy. The economy is usually healthy when it's acting like a riding tide that lifts all boats.
- AI is almost exclusively dominating the growth, with 13 of the 20 being AI oriented companies.
So, AI companies are in a boon phase. The rest are still struggling, and that's before you consider the wider economy beyond the stock market.
Hedgie@HedgieMarkets
🦔The S&P 500 closed at a record last Friday, but only 20 of its 500 members hit their own all-time highs. Of those 20, just seven had nothing to do with AI. Bank of America's Michael Hartnett pointed out that the exact same number of stocks hit new highs at the very top of the dot-com bubble in March 2000. The Nasdaq jumped 25% in April and May, its best two-month run in over 20 years, driven almost entirely by memory chip makers. Micron gained 88% in a single month. AMD 46%. SK Hynix 81%. Samsung 44%. Only about 55% of S&P 500 stocks sit above their 200-day moving average. My Take Every AI spending story I've covered on here, the circular financing, the $2-for-$1 burn rates, the fake token leaderboards, the $500 million Claude bills, this is where all of that meets the stock market. A record index price carried by 20 stocks is a market on a very small number of legs. The other 480 are not confirming the rally. BCA Research called it "extremely narrow" and Hartnett is advising clients to go defensive. I am not calling a crash and I'm not telling anyone to sell. But if your portfolio is concentrated in AI and semiconductors, you own the same trade as everyone else, and history says these narrow rallies either broaden out or they break. The dot-com comparison isn't perfect because Micron and AMD have real revenue at record volume. But Cisco had real revenue in 2001 too, and it didn't stop the multiple from compressing 80%. Revenue protects you from going to zero. It doesn't protect you from overpaying. Whether these prices already assume a future that may not arrive on schedule is something worth sitting with before the answer shows up in your brokerage account. Hedgie🤗
English
Third major AI-workstation announcement in like two months. mac studio ultra, dgx spark, now surface. local AI hardware went from niche to actually competitive really fast. curious to see if microsoft brings something different to the table once specs drop.
Microsoft Surface@surface
Introducing Surface Laptop Ultra. Built for world makers. Designed for what's next. The most powerful Surface laptop ever. Coming Fall 2026. Sign up to learn more: msft.it/6019vw79T
English
