Daniel Chernenkov

352 posts

Daniel Chernenkov banner
Daniel Chernenkov

Daniel Chernenkov

@danielckv

๐Ÿ“ธ Urban Explorer & Tech Entrepreneur โœจ 2x Post Exists. Staying Foolish, Building the Future. ๐ŸŒ CA / NY / TLV

California, United States Katฤฑlฤฑm Aralฤฑk 2013
50 Takip Edilen216 Takipรงiler
Daniel Chernenkov
Daniel Chernenkov@danielckvยท
Looking very well! Single-prompt training pipelines are excellent for rapid prototyping, but deterministic execution remains the core bottleneck. If the training set is small, a decoupled inference setup with a structured context vector store is almost always more cost-effective and controllable than running customized local weights. What did the VRAM footprint and training loss curve look like on your local setup?
English
0
0
1
115
Pietro Schirano
Pietro Schirano@skiranoยท
You can now vibe code a language model. From a single prompt, GPTโ€‘5.6 built the entire training pipeline and trained a model from scratch on my iMessage history. Locally on my Mac. It now generates replies in my writing style.
English
194
329
5.8K
702.2K
Daniel Chernenkov
Daniel Chernenkov@danielckvยท
3๏ธโƒฃ Alignment Padding: We align and pad your prompt boundaries to exact 1024-token limits to force upstream KV prompt cache hits every single time.
English
0
0
2
27
Daniel Chernenkov
Daniel Chernenkov@danielckvยท
Still paying full price to serve the exact same LLM queries? Swap one line of code to deploy Multi-Tier Semantic Caching and DataFission Intelligent Routing. Save 90%+ on your API bill or keep funding Sam's fusion reactor. betterproxy.ai
Daniel Chernenkov tweet media
English
6
1
72
62.9K
Daniel Chernenkov
Daniel Chernenkov@danielckvยท
2๏ธโƒฃ Seamless Rehydration: If the LLM doesn't explicitly fetch the crushed context, we automatically rehydrate the placeholders with your raw code back on the client stream.
English
0
0
2
32
Daniel Chernenkov
Daniel Chernenkov@danielckvยท
1๏ธโƒฃ Context Crusher: We instantly swap heavy code blocks and MCP payloads (>500 chars) with lightweight cryptographic placeholder tokens, hiding the original data behind an on-demand "fetch" tool definition.
English
0
0
2
36
Daniel Chernenkov retweetledi
Liad Yosef
Liad Yosef@liadyosefยท
Building AgentJourney by Ora. Release a swarm of real world top agents on your site and watch their live journeys - where they succeed, where they fail, and get insights on making your site agent-ready. Sneak peek: journey.ora.ai. Powered by @oradotai .
English
17
12
120
38.2K
Daniel Chernenkov
Daniel Chernenkov@danielckvยท
@alono88 ื–ื” ื’ื ืกื™ื™ื‘ืจ ืื‘ืœ.... ื™ืฉ ื‘ืขื™ื” ื’ื“ื•ืœื” ื™ื•ืชืจ ืฉื›ื‘ืจ ืขื•ืœื” ืขื›ืฉื™ื• ืžืฉืชื™ ื”ื—ื‘ืจื•ืช - ืžืฉืื‘ื™ื (+ืื ืจื’ื™ื”), ื•ื–ื” ืžื” ืฉืžืขื ื™ื™ืŸ ื‘ื”ื—ืœื˜ ื›ื™ ื›ื•ืœื ื• ื ื™ื–ื•ื ื™ื ืžื–ื”.
ืขื‘ืจื™ืช
1
0
1
64
Alon Oring
Alon Oring@alono88ยท
ืจื•ืฆื™ื ืœื”ืชืžื•ื“ื“ ืขื ื”ื‘ืขื™ื•ืช ื”ื’ื“ื•ืœื•ืช ื‘ื™ื•ืชืจ ืฉืžื˜ืจื™ื“ื•ืช ืืช OpenAI, ืื ืชืจื•ืคื™ืง, ื’ื•ื’ืœ, ืžื˜ื ื•ืืช ื›ืœ ื™ืชืจ ื”ื—ื‘ืจื•ืช ืฉืžืืžื ื•ืช ืžื•ื“ืœื™ ืฉืคื” ื’ื“ื•ืœื™ื? ืื– ืงื•ื“ื ื›ืœ ื›ื“ืื™ ืฉืชื‘ื•ืื• ืœื”ื’ื™ื“ ืฉืœื•ื ื‘ื›ื ืก CyberML ืฉืœ ืื•ืจื™ ื•ื“ื‘ืจ ืฉื ื™ ืื ื—ื ื• ืขื“ื™ื™ืŸ ืžื’ื™ื™ืกื™ื. ืงื™ืฉื•ืจ ื‘ืชื’ื•ื‘ื” ื”ืจืืฉื•ื ื”.
Uri Eliabayev@urieli17

ืื—ื“ ื”ื—ืœืงื™ื ื”ื›ื™ ืื”ื•ื‘ื™ื ืขืœื™ ื‘ืคืจืง ืขื ืขื•ืžืจ ืž Irregular ื”ื•ื ื”ื ืงื•ื“ื” ืฉื‘ื” ื”ื•ื ืžื“ื‘ืจ ืขืœ ืื™ืš ื—ื‘ืจื•ืช ืื—ืจื•ืช ืžื ืกื•ืช "ืœื’ื ื•ื‘" ืืช ื”ืžื•ื“ืœื™ื ืฉืœ ื—ื‘ืจื•ืช ื”-AI ื”ื’ื“ื•ืœื•ืช ื‘ืขื•ืœื. ืื ืชืคื ื• ืœืžื•ื“ืœ ืžืกืคื™ืง ืคืขืžื™ื ื“ืจืš ื”-API ืชื•ื›ืœื• ืœืฉื›ืคื•ืœ ืื•ืชื• ื‘ืฆื•ืจื” ืžืกืคื™ืง ื˜ื•ื‘ื”. ืžื™ ืฉื–ื•ื›ืจ, ื–ื• ื”ื™ื™ืชื” ื”ื˜ืขื ื” ืฉืœ ืื ื˜ืจื•ืคื™ืง ืœืคื ื™ ืžืกืคืจ ืฉื‘ื•ืขื•ืช ื›ื ื’ื“ ื—ื‘ืจื•ืช ื”-AI ื”ืกื™ื ื™ื•ืช. ืœื›ืื•ืจื”, ื”ื—ื‘ืจื•ืช ื”ืกื™ื ื™ื•ืช ืื™ืžื ื• ืืช ื”ืžื•ื“ืœ ืฉืœื”ื ืขืœ ืกืžืš ื“ืื˜ื” ืฉื”ื ืงื™ื‘ืœื• ืžื’ื™ืฉื” ื‘ืœืชื™ ืžืจื•ืกื ืช ืœืžื•ื“ืœื™ื ืฉืœ ืื ื˜ืจื•ืคื™ืง.

ืขื‘ืจื™ืช
7
1
56
9.4K
Dmytro G
Dmytro G@Dmitriy_Grey_AIยท
The architecture mismatch creates real challenges. Token economics fundamentally shift data pipeline priorities, requiring optimization for retrieval speed over storage. At RAI AI, we designed our infrastructure around this reality because customized analysis at scale demands rethinking data relevance per query.
English
1
0
1
16
Daniel Chernenkov
Daniel Chernenkov@danielckvยท
1/10 The LLM serving market is broken. Most teams are burning capital because they're treating Generative AI like a standard 2015 REST API. If you want to survive the shift to token-native infra, your mental model has to completely evolve. Letโ€™s talk architecture. ๐Ÿงต
Daniel Chernenkov tweet media
English
2
0
3
1.5K
Daniel Chernenkov
Daniel Chernenkov@danielckvยท
10/10 The economic building blocks of viable AI business models. Stop deploying models. Start engineering platforms.
English
0
0
0
24
Daniel Chernenkov
Daniel Chernenkov@danielckvยท
9/10 Thanks to Split-Prefill/Decode and prompt caching, the gateway seamlessly reroutes the context over to an active, on-demand node. The new node skips the heavy pre-fill penalty by leveraging the cached prompt matrix, picking up the decode phase right where it left off.
English
1
0
0
42
Daniel Chernenkov
Daniel Chernenkov@danielckvยท
ืื—ืœื” ืคื•ืกื˜. ืžื ื™ืกื™ื•ืŸ ืคื™ืชื•ื— ืฉืœื™ ืœืื—ืจื•ื ื” ื•ื”ืžื›ื™ืจื” ืœื ื‘ื™ื“ื™ื” - ืื ื™ ื™ื›ื•ืœ ืœื•ืžืจ ืฉื‘ืžืฆื‘ ื”ื ื•ื›ื—ื™ ืžื™ ืฉื—ื•ืฉื‘ ืขืœ ื“ืจืš ืฉืœ ืœืฉืœื•ื˜ ืขืœ ื›ืžื•ืช ืื“ื™ืจื” ืฉืœ ืฉืจืชื™ื ื–ื•ื›ื” ื‘ื˜ื•ื•ื— ื–ืžืŸ ื”ืžื™ื™ื“ื™ ืื‘ืœ ื™ืคื•ืœ ื‘ืขื•ื“ ืฉื ืชื™ื™ื ืœืื—ืจ ืฉื™ื‘ื™ืŸ ืฉืืคืฉืจ ืœืขืฉื•ืช ืขื™ื‘ื•ื“ ืฉื•ื ื” ื‘ืชื›ืœื™ืชื•. ื”ืจื•ื‘ ื‘ืขื™ืงืจ ื—ื•ืฉื‘ื™ื ืขืœ ืฉืจืช ืื—ื“ ืื‘ืœ ืคืจืกืžืชื™ ืคื•ืกื˜ ืœื ืžื–ืžืŸ ืขืœ ืžื—ืฉื‘ื” ืฉืœ Token as a Service
ืขื‘ืจื™ืช
0
0
0
52
Alon Oring
Alon Oring@alono88ยท
ืื– ื›ืžื” ืขื•ืœื” ืœื”ืจื™ืฅ ืžื•ื“ืœื™ ืฉืคื”? ืขื ื”ื™ืฆื™ืื” ืฉืœ ืžื•ื“ืœื™ื ื‘ืงื•ื“ ืคืชื•ื—, ืืคืฉืจ ืœืงื‘ืœ ื”ืขืจื›ื” ื’ืกื” ืœื›ืžื” ืžืฉืื‘ื™ ื—ื™ืฉื•ื‘ ื ื“ืจืฉื™ื ืขืœ ืžื ืช ืœื”ืจื™ืฅ "ืžืงื•ืžื™ืช" ืžื•ื“ืœื™ ืฉืคื” ื’ื“ื•ืœื™ื ืฉืžืกื•ื’ืœื™ื ืœื”ืชื—ืจื•ืช ื‘ื‘ื™ืฆื•ืขื™ื ืฉืœ ื”ื—ื‘ืจื•ืช ื”ื’ื“ื•ืœื•ืช. ื ื™ืงื— ืœื“ื•ื’ืžื” ืืช DeepSeek 4 Pro, ืžื•ื“ืœ ื‘ื’ื•ื“ืœ 1.6 ื˜ืจื™ืœื™ื•ืŸ ืคืจืžื˜ืจื™ื ืฉืžืชืงืจื‘ ื‘ื’ื•ื“ืœ ืฉืœื•, ืœืคื™ ืคืจืกื•ืžื™ื ืฉืœ ืื™ืœื•ืŸ ืžืืกืง, ืœ-5 ื˜ืจื™ืœื™ื•ืŸ ืคืจืžื˜ืจื™ื ืฉืœ ืžื•ื“ืœื™ื ื›ืžื• Opus. ืจืืฉื™ืช, ืขืœ ืžื ืช ืœืื›ืœืก ืืช ื“ื™ืคืกื™ืง 4 ืคืจื• ื‘ื–ื™ื›ืจื•ืŸ, ืื ื—ื ื• ืฆืจื™ื›ื™ื ื›-900GB ืฉืœ ืจืื. ื”ื“ืจืš ื”ื›ื™ ื™ืขื™ืœื” ืœื”ืฉื™ื’ ืืช ื”ื›ืžื•ืช ื”ื–ื• ื”ื™ื ื‘ืืžืฆืขื•ืช ืฉืžื•ื ื” ื›ืจื˜ื™ืกื™ ืžืกืš ืžืกื•ื’ H200. ื›ืœ ื›ืจื˜ื™ืก ื›ื–ื” ืขื•ืœื” ื›-30,000 ื“ื•ืœืจ ื•ืขืœ ืžื ืช ืœื—ื‘ืจ ืื•ืชื ืื—ื“ ืœืฉื ื™ ื ืฆื˜ืจืš ืฉืจืช ืฉื™ืขืœื”, ื™ื—ื“ ืขื ื”ื›ืจื˜ื™ืกื™ื ืขืฆืžื, ื‘ืขืจืš 400 ืืœืฃ ื“ื•ืœืจ (ืœื ื›ื•ืœืœ ื—ืฉืžืœ, ืงื™ืจื•ืจ ื•ืชื—ื–ื•ืงื”). ืœื ื ืขื™ื. ืžื” ื‘ื ื•ื’ืข ืœืขื ืŸ? ืฉืจืช ื–ื”ื” ื™ืขืœื” ืœื ื• ื›-100 ื“ื•ืœืจ ืœืฉืขืช ืจื™ืฆื” ืื• 72,000 ื“ื•ืœืจ ืœื—ื•ื“ืฉ, ื‘ื”ื ื—ื” ื•ื”ืฉืจืช ืจืฅ ืœืœื ื”ืคืกืงื”. ืื ื™ ืžืชืขืœื ืœืจื’ืข ืžื”ืขื•ื‘ื“ื” ืฉืœื ื ื™ืชืŸ ืœืงื‘ืœ ืืช ื”ืฉืจืชื™ื ื”ืœืœื• ืžืกืคืงื™ื•ืช ื”ืขื ืŸ ื‘ืื•ืชื” ืงืœื•ืช ื‘ื” ืืคืฉืจ ืœืงื‘ืœ ืฉืจืชื™ื ืคืฉื•ื˜ื™ื ื™ื•ืชืจ. ื‘ื›ืžื” ืžืชื›ื ืชื™ื ืžืกื•ื’ืœ ืœืชืžื•ืš ืฉืจืช ื›ื–ื”? ืœื ื”ืจื‘ื”. ืฉืจืช ื›ื–ื” ืžืกื•ื’ืœ ืœืกืคืง ื›ืžื” ืืœืคื™ ื˜ื•ืงื ื™ื ื‘ืฉื ื™ื” ื‘ืฉื™ื, ื›ืืฉืจ ืื ื—ื ื• ืžืฉืชืžืฉื™ื ื‘ืกื•ื›ื ื™ ืงื™ื“ื•ื“ ื›ืžื• Claude Code ื•ื”ืื•ืจืš ืฉืœ ื”ืกืฉืŸ ื’ื“ืœ, ืคืจื•ืžืคื˜ ื™ื—ื™ื“ ื™ื›ื•ืœ ืœื”ื›ื™ืœ ืขืฉืจื•ืช ื•ืžืื•ืช ืืœืคื™ ื˜ื•ืงื ื™ื. ื›ืžื•ื‘ืŸ ืฉืœื ื›ืœ ื”ืžืคืชื—ื™ื ืขื•ื‘ื“ื™ื ื‘ืžืงื‘ื™ืœ ื‘ื›ืœ ืจื’ืข ื ืชื•ืŸ (ืื ื™ ืžืงื•ื•ื” ืฉืืชื ืงื•ืจืื™ื ืืช ื”ืงื•ื“ ืœืคื ื™ ืฉืืชื ืœื•ื—ืฆื™ื ืื ื˜ืจ) ืื– ืืคืฉืจ ืœื”ืขืจื™ืš ืฉื”ื—ืกื ื”ืขืœื™ื•ืŸ ืœืงื™ื“ื•ื“ ืืคืงื˜ื™ื‘ื™ ื”ื•ื ื›-20 ืžืคืชื—ื™ื ืขืœ ืฉืจืช ืื—ื“. ืœื•ื ื’ ืฉื‘ื‘ื™ื (ื™ืฉ ืขื ื™ื™ืŸ, ืœื ื”ืžืœืฆื”).
ืขื‘ืจื™ืช
35
4
241
30.4K
Daniel Chernenkov
Daniel Chernenkov@danielckvยท
Andrew, your perspective is intriguing. However, I believe coding speed is a superficial metric. The true bottleneck lies not in syntax generation but in managing state, data lineage, and system-wide integrity. Agents that prioritize code generation without a deep understanding of architecture are merely accelerating technical debt.
English
0
0
0
115
Andrew Ng
Andrew Ng@AndrewYNgยท
Coding agents are accelerating different types of software work to different degrees. When we architect teams, understanding these distinctions helps us to have realistic expectations. Listing functions from most accelerated to least, my order is: frontend development, backend, infrastructure, and research. Frontend development โ€” say, building a web page to serve descriptions of products for an ecommerce site โ€” is dramatically sped up because coding agents are fluent in popular frontend languages like TypeScript and JavaScript and frameworks like React and Angular. Additionally, by examining what they have built by operating a web browser, coding agents are now very good at closing the loop and iterating on their own implementations. Granted, LLMs today are still weak at visual design, but given a design (or if a polished design isnโ€™t important), the implementation is fast! Backend development โ€” say, building APIs to respond to queries requesting product data โ€” is harder. It takes more work by human developers to steer modern models to think through corner cases that might lead to subtle bugs or security flaws. Further, a backend bug can lead to non-intuitive downstream effects like a corrupted database that occasionally returns incorrect results, which can be harder to debug than a typical frontend bug. Finally, although database migrations can be easier with coding agents, theyโ€™re still hard and need to be handled carefully to prevent data loss. While backend development is much faster with coding agents, they accelerate it less, and skilled developers still design and implement far better backends than inexperienced ones who use coding agents. Infrastructure. Agents are even less effective in tasks like scaling an ecommerce site to 10K active uses while maintaining 99.99% reliability. LLMs' knowledge is still relatively limited with respect to infrastructure and the complex tradeoffs good engineers must make, so I rarely trust them for critical infra decisions. Building good infrastructure often requires a period of testing and experimentation, and coding agents can help with that, but ultimately thatโ€™s a significant bottleneck where fast AI coding does not help much. Lastly, finding infrastructure bugs โ€” say, a subtle network misconfiguration โ€” can be incredibly difficult and requires deep engineering expertise. Thus, Iโ€™ve found that coding agents accelerate critical infrastructure even less than backend development. Research. Coding agents accelerate research work even less. Research involves thinking through new ideas, formulating hypotheses, running experiments, interpreting them to potentially modify the hypotheses, and iterating until we reach conclusions. Coding agents can speed up the pace at which we can write research code. (I also use coding agents to help me orchestrate and keep track of experiments, which makes it easier for a single researcher to manage more experiments.) But there is a lot of work in research other than coding, and todayโ€™s agents help with research only marginally. Categorizing software work into frontend, backend, infra, and research is an extreme simplification, but having a simple mental model for how much different tasks have sped up has been useful for how I organize software teams. For example, I now ask front-end teams to implement products dramatically faster than a year ago, but my expectations for research teams have not shifted nearly as much. I am fascinated by how to organize software teams to use coding agents to achieve speed, and will keep sharing my findings in future posts. [Original text: deeplearning.ai/the-batch/issuโ€ฆ ]
English
85
106
559
74.4K