34,950 results · Most relevant
Nouha Dziri endorses replacing Sol sub-workers with cheap Luna workers in the planning-heavy setup, which yields almost the same performance at 60% less cost, losing only a little performance. A strong lead with cheap workers (Sol + Luna) achieves the same performance as Sol + Sol but at much lower cost.
Nouha Dziri recommends splitting the work across a team: a lead keeps the task and hands out subtasks, and each worker starts with a fresh context. Each worker can read files, run tests, and dig through logs, then send back only a compact result, so the lead keeps the full picture and can finish faster.
To optimize the agent setup for good reward at much cheaper cost, Nouha Dziri recommends taking the same lead and giving it workers at only one stage of the software engineering process.
The narrator advises owning the page that an AI Overview quotes, because that link sits above every blue result. The new job in AI search is getting named and getting cited: if you get named in an AI answer's short list, you've been recommended to a buyer who hasn't even clicked yet.
To get started with the Figma GitHub app, connect your Figma organization to your GitHub organization or account, select the GitHub account or organization to connect, and choose whether to authorize all repositories or a select set. The first few steps are the same for all Figma admins, and the repository authorization choice can be changed later.
A step-by-step walkthrough for connecting a Figma organization to GitHub: open Admin from the file browser, go to Settings > Resources, select Figma GitHub App, click Connect GitHub, and install the app from the Connect to GitHub modal.
After installation you can manage which repositories the Figma app can see: from your GitHub account, navigate to settings, click GitHub Apps, select Configure next to the Figma app, and under Repository Access choose All Repositories or select repositories, then click Save. To use repositories with code layers in Figma Make, you must allow cloning of repositories.
If running Blackwell-generation GPUs, don't run anything higher than FP4 quantization; FP4 is best, and going from FP8 to FP4 saves roughly 40% of GPU footprint for zero loss in precision. Long context is a scam: even though you can serve one million tokens, 256K is really what you need. Consider tensor parallelism: even though GLM fits on TP4, you may not have enough KV cache, so TP8 might work better depending on workload size. Coding tasks need to be above 90% cache hit, and you can look into optimized speculative decoding, sequential pre-fill, and PDD disaggregation as esoteric tuning options.
TP8 might work better depending on workload size
Set rules for operations you don't want to allow (e.g., allow grep but not gradle); this gives regulators confidence the system is not an autonomous agent needing different supervision.
If starting from scratch, Aaron Levie would look at the industry/job-function matrix, pick an industry he knows extremely well, reverse engineer what the work looks like in that industry, and build an agent for it. The second best scenario is having a co-founder that has done the job you want to automate; the third best is to go find a way to do the job yourself, like shadowing somebody and doing it all day long.
attributed to Levie as his personal approach ('if I had to literally start from scratch')
SGLang's president describes how the team manages memory pools for hybrid architectures: they allocate one shared pool of physical memory without per-pool-type limits, letting pools grow dynamically from opposite ends depending on context length. They build a unified radix tree on full-token IDs across GPU memory, host memory, and external storage as L1, L2, and L3 cache tiers. They introduced a compaction operation that reorganizes memory fragments in physical memory space using low-latency memory copy kernels and remaps virtual page IDs by only changing the low-level mapping. The president explains that managing memory pools separately creates major challenges for efficient cache sharing, and that using separate standalone radix trees per memory pool component brings compatibility and
Founders should hire a filmmaker as one of their first hires — it was his number one planned hire even before starting the company (along with an engineer). The reason: no founder wants to put daily attention on making videos. At his last company the film team was two people on a seven-person team; at Clicky at one point three people made videos and there was one engineer.
GaryVee's concrete playbook: if you like jigsaw puzzles, film yourself doing the puzzle over a beer with Love Island in the background and post clips on TikTok. The content is the activity you'd already be doing, so it only works if it's genuinely what you want to do. Consistency and perseverance, plus learning a little something, are prerequisites.
illustrative scenario, not verified outcomes
Keep the input form visible and expanded directly on the landing page so people can type right into it, because conversion was much higher than requiring a click-through, popup, or email gate. Removing the visible prompt input box from their landing page caused their conversion rate to totally tank.
self-reported anecdote, no numbers given
Isabelle Williams describes a flow where an agent takes anonymous actions in your app with limited capabilities, receiving an access token plus a claim token. When a user wants to take ownership of what their agent did, the agent gives them a link and a one-time code to enter at sign-up; the user signs up, signs in, enters the code, and claims the agent's work under their account. She notes that requiring a signup flow with link, code entry, and email verification is where you lose your user in product-led growth.
Kurtz advises auditing the organization for 'Shadow AI' — finding where AI agents are operating, who is doing what, and under what identity — since nobody knows where rogue AI is. He describes CrowdStrike's approach of instrumenting agents from birth to capture the entire chain of events (sub-agents, tool calls, MCP server access, data pulls) and then adding policy and blocking rules around them. He also warns against unleashing agents with broad identities and unrestricted package-downloading permissions, which creates tremendous exposure.
Brockman urges defenders to use peak frontier AI capabilities during the defender's window to find and fix vulnerabilities in their systems, applying each new model to their infrastructure until it saturates, then repeating with the next model. OpenAI's Daybreak program is cited as an example of responsibly giving defenders access.
Sam McKenna recommends asking the prospect who else would benefit from this in your organization, and whether peers on other teams or in other lines of business would need to see this before deciding to move forward, since prospects often have pals with the same issues who could be brought onto the call.
When a prospect accepts a meeting, follow three steps: respond with gratitude before booking, send a calendar invite with your name first and a clear purpose in the subject line, and connect with the buyer on LinkedIn with a thank-you message.
The speaker advises letting UGC creators submit statics, requiring them to prove themselves first, and having video-script creators write statics too since they already craft the perfect script in their minds.
Choose sources → to get a library of your own and a daily brief.