Neil Lock

BREAKING: 86.8% SWE-bench Verified, Biggest Leap Ever. "SWE Bench verified has been the Super Bowl of AI coding benchmarks for the past two years. The last time the coding benchmarks had a double digit leap was March of 2024. This ushered in a huge era for AI coding. The AI agent was here." -Brian Elliott We just proved the AI pundits wrong at Google's Sunnyvale campus. While everyone's saying "AI acceleration is plateauing**,** we took a different approach at Blitzy. The results? 86.8% on SWE-Bench verified - the industry's premier AI coding benchmark that tests systems on real-world GitHub issues - delivering a 10% lift over the current leader and the first double-digit leap since March 2024. Here's what changed the game: --> System 2 AI: Trading time for quality. Our platform spends hours mapping every line of code relationally and hierarchically summarizing entire codebases (100M+ lines). --> Multi-agent orchestration: Extended inference time compute with validation across multiple models with built-in prompt engineering. As I told the audience: "The next frontier of progress is not just pre-training models. By driving more inference time compute into your systems, you are able to drive up the quality of output dramatically." Link to white paper in comments. #blitzy #ainativesdlc #autonomoussoftwaredevelopment #aicoding #nextfrontier

262

39 Comments

Most teams prompt and pray. The best teams delegate and scale. Prompting was the frontier. Now it's autonomy. Agents > Tools > RAG > LLMs. Each layer adds memory, action, autonomy. 🔷 LLMs were closed-book systems. Prompt in; answer out. No context. No awareness. 🔷 RAG gave models a working memory. It retrieved facts in real time, not just recall. 🔷 RAG is architecture: not just an add-on. Embeddings and chunking shape its value. 🔷 Tools made models interact with systems. They used APIs, files, and databases. 🔷 Orchestration made tools reliable. Chaining, routing, recovery: it's glue code. 🔷 Agents pursue goals, not just answers. They plan, retry, adapt, and improve. 🔷 Autonomy still needs constraints. Memory, oversight, and sync keep it useful. 🔷 Multi-agent setups unlock scale. Agents act as writers, analysts, and more. Workflows have changed how things get done. They are now built for agents, not people. Are you still prompting or building agents? ♻️ Repost to help someone learn AI the right way ➕ Follow me, Sairam, for AI engineering that works ----- Want to go deeper? Gradient Ascent has you covered. Join 21k+ readers from Google, Meta, Netflix, and over 154+ countries worldwide.

278

91 Comments

The Meta Product Manager Mafia has raised more than $110M in funding and comprises 20+ companies. Product managers at Meta have expectations and responsibilities that make them great entrepreneurs: 1) Obsess over the North Star: Meta PMs are trained to align cross-functional teams toward a single, company-defining goal. They’re responsible for maintaining clarity and direction. 2) Simplifying complex problems: This is a rare skill but an important one for any PM. They take complex problems and break them down into clean processes that designers, engineers, and execs can all understand and rally behind. 3) Experiment and fail quickly: Meta PMs are encouraged to ship ideas quickly, test them in-market and if it fails the teams are broken down aggressively to focus on something else. 4) Brutal transparency: As a Meta PM, there’s nowhere to hide. Frequent exec reviews mean PMs are directly accountable to VPs and C-suite. The expectations for growth and comp are communicated candidly and that visibility makes great PMs rise quickly.  Here are the top 10 most funded former Meta PMs, turned entrepreneurs: 1. Archy: Co-founded by Jonathan Rat Funding Raised: $27M 2. Lendtable: Co-founded by Mitchell Jones Funding Raised: $22.5M 3. Greenlite AI: Co-founded by Will Lawrence Funding Raised: $20M 4. Things, Inc. / Rooms: Co-founded by Jason Toff Funding Raised: $11M 5. LastMile AI: Co-founded by Andrew Hoh Funding Raised: $10M 6. Kopperfield: Co-founded by Mark Wong Funding Raised: $5M 7. Ribbon - Inheritance Platform: Co-founded by Saeid Kian Funding Raised: $2.7M 8. GovEagle: Co-founded by Akash M. Funding Raised: $2.5M 9. Rally (YC W25): Founded by Thiago Peres Funding Raised: $500K 10. Besimple AI: Co-founded by Yi Zhong Funding Raised: $500k Data sourced from Crustdata (YC F24) Did we miss out on anyone?

362

36 Comments

🚀 OneRec vs. OneSearch: Same "one-stage" vision, different token strategies A big trend in recommender & search systems is collapsing retrieval + ranking into a single generative stage. Two recent papers (both from Kuaishou) show how this plays out differently for recommendations vs. search: 🔹 OneRec (recommendation) - Items are tokenized into semantic IDs (via multimodal + collaborative quantization). - The model directly generates item IDs from user behavior sequences. - At inference: no natural language tokens are involved — everything happens in the semantic ID space. 🔹 OneSearch (search) - Queries still use language tokens as input → crucial for capturing keyword intent. - Items are discretized into keyword-enhanced semantic IDs (KHQE). - The model maps query text tokens → SID tokens. - At inference: it outputs item IDs, but only after fully understanding the query in natural language. ⚖️ Why this matters - Recs are query-less → the system leans entirely on learned item embeddings & user preference. - Search is query-anchored → language tokens must stay in the loop to preserve lexical fidelity before handing off to semantic IDs. ✨ Takeaway: Both unify retrieval & ranking, but OneRec drops language at inference, while OneSearch keeps it on the input side to ensure query relevance.

54

12 Comments

How Stripe prevents double payments (explained in 2 mins or less): 1 They use idempotent APIs to prevent double payment. 2 They use UUID as the idempotency key. 3 They send the idempotency key through HTTP headers. 4 They create a new idempotency key whenever the request payload changes. 5 They store the idempotency keys in a database on the server side. 6 They query the database with keys to check if a request was processed earlier. 7 They process a request only if it's new and then store its idempotency key in the database. 8 They roll back a transaction using the ACID database on a server error. 9 They remove the idempotency keys from the database after 24 hours for reuse. 10 They use exponential back-off with jitter to avoid the thundering herd problem. What else would you add? —— 👋 PS - If you enjoyed this post, you'll love my detailed article about this case study (with visuals): → https://lnkd.in/erMkqwq4 ——— ♻ Repost to help others learn idempotent APIs. 👤 Follow Neo Kim for more.

2,546

139 Comments

I was going through a report from Menlo Ventures released on Thursday. Anthropic just overtook OpenAI in enterprise usage: 32% vs 25% market share. That's a complete reversal from 2023 when OpenAI dominated with 50%. Anthropic jumped from $1B to $4B in revenue in just 6 months, while targeting a smaller, higher-value enterprise segment. Both companies' flagship models are essentially at parity in 2025. It's about Claude's focus on logic, structure, and regulatory compliance, priorities that enterprise buyers actually pay premium for. AI spending is exploding 44% year-over-year, hitting $371B in 2025. And, nearly half of all AI spending is shifting from training to inference by 2032. While everyone's obsessing over model benchmarks, you should push your team to focus on: ➤ Code generation (already a $1.9B market) ➤ Agent-first architectures ➤ Production inference optimization ➤ Enterprise tool integration Model costs dropped 280x from 2022 to 2024, but enterprise spending keeps growing. Why? Because enterprises pay for outcomes, not tokens. They want AI that integrates with their systems, understands their data, and solves their specific problems. What I can say from all of this: stop chasing the latest model release. Start building AI infrastructure that enterprises actually want to buy. The companies winning aren't the ones with the best models. They're the ones solving real business problems at scale. #AI #Consulting #Strategy #Startups

104

28 Comments

The hottest role in AI startups right now isn’t Forward Deployed Engineers. It isn't GTM Engineers. It’s Deployment Strategists. Decagon calls it an “Agent Product Manager.” Harvey calls it a “Solutions Architect.” Palantir Technologies has had versions of this role for years. And the salaries are climbing fast: - Decagon: $200k–$285k - Palantir Technologies: $120k–$200k - Figma: $150k–$260k - Ramp: $100k–$180k - Harvey: $190k–$260k So who are these people? They are usually pseudo-technical -- CS or engineering majors, or folks with technical work experience. Many come from 2 years in consulting, IB, or PE, then jump into startups to get their hands dirty. They are young, hungry, polished, and comfortable being in front of customers. What do they actually do? They make sure enterprise AI deployments succeed. A $100k+ deal does not survive on a nice pitch or a self-serve onboarding flow. It survives if the customer sees value in the pilot. That means: - Embedding directly with the customer - Designing prompt logic for specific workflows - Working with engineering to align integrations and data flow - Helping exec teams define their AI roadmap - Running feedback loops into product and GTM Why does this role matter so much? Because enterprise AI is messy. Integrations, data transfer, and adoption make or break a deal. Most buyers are using AI for the first time, and each has unique workflows. Deployment Strategists bridge that gap. They own the outcome. They are accountable for making pilots successful, which often means millions in revenue down the line. At Conversion, Sam Bochner has been leading this work for us. We are now thinking about scaling it into a full team. Because a few successful pilots can fund an entire department, and the cost of failed deployments is too high to ignore. Is this just a rebrand of customer success? Not really. Success is about answering tickets and renewals. Deployment Strategy is about going deep with a few enterprise accounts, extracting maximum value, and ensuring the pilot closes into a multi-year contract. Call it Agent PM, Solutions Architect, or Deployment Strategist. Whatever the title, this is becoming one of the most important roles in AI SaaS.

643

90 Comments