/Orbit
EVERYTHING YOU FOLLOW - DISTILLED
MONDAY · 3 AUG 2026
806 TRACKED · 23 POSTED · 36 ITEMS
From your YouTube · ranked
19
CHANNELS
POSTED
POSTED
28
VIDEOS
ANALYZED
ANALYZED
AEAI Engineer ×5SOStanford Online ×5FForbes ×2CPCraig Percoco ×1RWRob Walling ×1WRWes Roth ×1PYPeter Yang ×1TAThe AI Automators ×1NMNew Money ×1TKTK Kader ×1TTED ×1ASAustin Schneider ×1SSStarter Story ×1ANAI News & Strategy Daily ×1LSLenny's Podcast ×1VVeritasium ×1CUChina Uncensored ×1RCReventure Consulting ×1IInspirationTuts ×1
01
0:01In May 2022, US authorities seized the Amadea, a 348-foot superyacht, in Fiji under Task Force KleptoCapture as a DOJ action.5:43Sajwani targets only unrepeatable locations—Sheikh Zayed Road, Dubai Water Canal, Bulgari, The Palm—as his core competitive moat.10:48The development includes a 200,000 sq ft spa, 200,000 sq ft food hall, two members clubs, and more than a dozen restaurants.
0:01In May 2022, US authorities seized the Amadea, a 348-foot superyacht, in Fiji under Task Force KleptoCapture as a DOJ action.5:43Sajwani targets only unrepeatable locations—Sheikh Zayed Road, Dubai Water Canal, Bulgari, The Palm—as his core competitive moat.10:48The development includes a 200,000 sq ft spa, 200,000 sq ft food hall, two members clubs, and more than a dozen restaurants.02
0:12Post-training—teaching models to keep improving during real-world deployment—is the next major frontier in AI, per Applied Compute's Raymond Feng.8:21Poor environment design causes RL-trained models to degrade into gibberish over time; the environment is the single biggest failure point.15:34Any task a model can be improved on lets it be treated as one deployment that learns across many simultaneous interactions in production.
0:12Post-training—teaching models to keep improving during real-world deployment—is the next major frontier in AI, per Applied Compute's Raymond Feng.8:21Poor environment design causes RL-trained models to degrade into gibberish over time; the environment is the single biggest failure point.15:34Any task a model can be improved on lets it be treated as one deployment that learns across many simultaneous interactions in production.03
0:12General Reasoning, a London-based RL company led by CEO Ross Taylor, is working on extending reinforcement learning to long time horizons.8:15Their failed experiments converged on the Bitter Lesson: scale and raw compute beat custom engineering tricks every time.15:24They traded off-policy correctness for GPU utilization, but as they scaled further, a new issue emerged that this trade-off couldn't absorb.
0:12General Reasoning, a London-based RL company led by CEO Ross Taylor, is working on extending reinforcement learning to long time horizons.8:15Their failed experiments converged on the Bitter Lesson: scale and raw compute beat custom engineering tricks every time.15:24They traded off-policy correctness for GPU utilization, but as they scaled further, a new issue emerged that this trade-off couldn't absorb.04
05
0:00Reaching $1M ARR is hard, but the $1M-to-$10M gap breaks founders in an entirely different way—most never cross it.8:21The second critical system is accountability, which requires a weekly operating rhythm to keep clients and teams on track.15:43GTM is the top priority; SEO, which many SaaS companies still depend on heavily, is declining and an unreliable growth lever.
0:00Reaching $1M ARR is hard, but the $1M-to-$10M gap breaks founders in an entirely different way—most never cross it.8:21The second critical system is accountability, which requires a weekly operating rhythm to keep clients and teams on track.15:43GTM is the top priority; SEO, which many SaaS companies still depend on heavily, is declining and an unreliable growth lever.06
0:00Noam Brown of OpenAI posted a result the speaker calls front-page news: an AI breakthrough on mathematics that went largely unnoticed.11:14Every prior result tested appeared imitable and approximatable, but the Astra model broke that pattern and couldn't be explained away.21:12The unresolved challenge ahead: how to balance AI's ability to solve disease and scarcity against the risks that come with that power.
0:00Noam Brown of OpenAI posted a result the speaker calls front-page news: an AI breakthrough on mathematics that went largely unnoticed.11:14Every prior result tested appeared imitable and approximatable, but the Astra model broke that pattern and couldn't be explained away.21:12The unresolved challenge ahead: how to balance AI's ability to solve disease and scarcity against the risks that come with that power.07
0:00Open-source winning is the stated core mission: the goal is to give everyone freedom over AI, not to build closed proprietary systems.21:01Hermes's tool gateway removes the need to separately sign up for additional tools—they're accessible directly from within the portal.39:44The commitment is to keep distributing top-tier AI intelligence to everyone equally, so no individual or company gains an unfair advantage.
0:00Open-source winning is the stated core mission: the goal is to give everyone freedom over AI, not to build closed proprietary systems.21:01Hermes's tool gateway removes the need to separately sign up for additional tools—they're accessible directly from within the portal.39:44The commitment is to keep distributing top-tier AI intelligence to everyone equally, so no individual or company gains an unfair advantage.08
0:00Serious users of Claude Code, Pi Agent, or Hermes Agent need an observability layer — Langfuse is positioned as that essential free tool.12:17System prompts grow large enough over time to develop internal contradictions, and that is when model behavior starts to degrade.23:12Redacting sensitive information is mandatory with observability platforms, and Langfuse demonstrates this as a built-in requirement.
0:00Serious users of Claude Code, Pi Agent, or Hermes Agent need an observability layer — Langfuse is positioned as that essential free tool.12:17System prompts grow large enough over time to develop internal contradictions, and that is when model behavior starts to degrade.23:12Redacting sensitive information is mandatory with observability platforms, and Langfuse demonstrates this as a built-in requirement.09
0:00Pabrai frames the current investing environment as a gold rush; his book promises a low-risk value method to high returns.22:22Free investment ideas are available at valueinvestorsclub.com — submit your email and access them at no cost.42:14Management honesty and capability are non-negotiable filters — any investment that fails either test is disqualified.
0:00Pabrai frames the current investing environment as a gold rush; his book promises a low-risk value method to high returns.22:22Free investment ideas are available at valueinvestorsclub.com — submit your email and access them at no cost.42:14Management honesty and capability are non-negotiable filters — any investment that fails either test is disqualified.10
0:00This guide targets non-Stanford founders who cannot raise VC pre-seed money and must build a real SaaS from $0 to first revenue.15:23Vibe-coding now lets you ship before charging, but the real test is unchanged: are people paying, or just being polite.29:03The second strategy introduced is the AI monetization digital twin — a pricing construct tied to how users engage with AI features.
0:00This guide targets non-Stanford founders who cannot raise VC pre-seed money and must build a real SaaS from $0 to first revenue.15:23Vibe-coding now lets you ship before charging, but the real test is unchanged: are people paying, or just being polite.29:03The second strategy introduced is the AI monetization digital twin — a pricing construct tied to how users engage with AI features.11
0:03The devastated Altadena neighborhoods were defined by total silence — no ambient sound survived, making absence itself the loudest marker of destruction.4:48Her best friend died of cancer there a year ago; that friend's partner fled the fires with only a small suitcase and a loaf of bread.8:58A fragile piece of paper survived a lithium battery fire so intense that even firefighters were stunned by its endurance.
0:03The devastated Altadena neighborhoods were defined by total silence — no ambient sound survived, making absence itself the loudest marker of destruction.4:48Her best friend died of cancer there a year ago; that friend's partner fled the fires with only a small suitcase and a loaf of bread.8:58A fragile piece of paper survived a lithium battery fire so intense that even firefighters were stunned by its endurance.12
0:00Working late into the night to ensure no detail is missed is the price of scaling the agency — completeness over comfort.25:59Scaling the agency requires not relying on any single element — both speakers agree on this as a firm principle.49:06Even as the agency scales, the owner insists on staying hands-on — attending shoots and personally running weekly client calls.
0:00Working late into the night to ensure no detail is missed is the price of scaling the agency — completeness over comfort.25:59Scaling the agency requires not relying on any single element — both speakers agree on this as a firm principle.49:06Even as the agency scales, the owner insists on staying hands-on — attending shoots and personally running weekly client calls.13
0:00Weekly hours vary wildly — sometimes just 1-2 hours a day, sometimes 8-10 — and every week's output is genuinely different.4:23After persisting through a rough stretch, the app hit $15K/month at its peak and has snowballed since toward the $40K/month run rate.8:15Sleep is non-negotiable: 8 hours every night regardless of the temptation to stay up, because it directly sustains productive output.
0:00Weekly hours vary wildly — sometimes just 1-2 hours a day, sometimes 8-10 — and every week's output is genuinely different.4:23After persisting through a rough stretch, the app hit $15K/month at its peak and has snowballed since toward the $40K/month run rate.8:15Sleep is non-negotiable: 8 hours every night regardless of the temptation to stay up, because it directly sustains productive output.14
0:00Builders discouraged by OpenAI and Anthropic launches are stuck at Level 1 — reacting to surface releases instead of building on deeper understanding.6:29The new requirement for AI builders is deep domain-specific understanding of AI, not just general familiarity with the tools.12:13The strongest builders anticipate AI's trajectory and use it as a tailwind in their domain, turning every model release into a competitive advantage.
0:00Builders discouraged by OpenAI and Anthropic launches are stuck at Level 1 — reacting to surface releases instead of building on deeper understanding.6:29The new requirement for AI builders is deep domain-specific understanding of AI, not just general familiarity with the tools.12:13The strongest builders anticipate AI's trajectory and use it as a tailwind in their domain, turning every model release into a competitive advantage.15
0:00Scaling created a rigid pod ratio — one designer and PM per six engineers — that became an unexamined default constraining how product orgs operate.38:14Video artifacts of real user sessions let teams distinguish a comprehension gap (user misunderstands the product) from a product gap (it doesn't work as expected).1:12:10A structural change leadership feared would damage team culture was actually implemented; the outcome proved the concern unfounded, building confidence for bolder moves.
0:00Scaling created a rigid pod ratio — one designer and PM per six engineers — that became an unexamined default constraining how product orgs operate.38:14Video artifacts of real user sessions let teams distinguish a comprehension gap (user misunderstands the product) from a product gap (it doesn't work as expected).1:12:10A structural change leadership feared would damage team culture was actually implemented; the outcome proved the concern unfounded, building confidence for bolder moves.16
0:00The test uses the largest spider silk sample ever assembled, making it the most rigorous strength comparison yet attempted.12:07C-terminal ends link neighboring spidroins together while N-terminal ends stay mostly separate, keeping the proteins loose enough to flow before being drawn into fiber.22:57What looks unremarkable is revealed to be a single unbroken continuous filament — the entire thread is one strand.
0:00The test uses the largest spider silk sample ever assembled, making it the most rigorous strength comparison yet attempted.12:07C-terminal ends link neighboring spidroins together while N-terminal ends stay mostly separate, keeping the proteins loose enough to flow before being drawn into fiber.22:57What looks unremarkable is revealed to be a single unbroken continuous filament — the entire thread is one strand.17
0:12AI's benchmark obsession follows a hype-cycle pattern — chasing leaderboard scores substitutes for defining what good AI behavior actually means in a product context.7:48Eval design is an expression of product values — what you want AI to do and how to behave — so building useful evals requires product sense, not just engineering.14:33Good evals demand full bidirectional coverage: every prompt requirement must have a verifier, and every verifier must map to something the prompt actually asks for.
0:12AI's benchmark obsession follows a hype-cycle pattern — chasing leaderboard scores substitutes for defining what good AI behavior actually means in a product context.7:48Eval design is an expression of product values — what you want AI to do and how to behave — so building useful evals requires product sense, not just engineering.14:33Good evals demand full bidirectional coverage: every prompt requirement must have a verifier, and every verifier must map to something the prompt actually asks for.18
0:00The CCP has located where wealthy citizens are hiding money and is moving to seize it; separately, microchip smugglers are being caught importing banned chips into China.5:19The same restricted materials are also flowing to Russia, exposing a double standard that undermines the credibility of enforcement efforts.10:06China directly subsidizes the chemical companies supplying fentanyl precursors, structurally contradicting any meaningful cooperation with US enforcement.
0:00The CCP has located where wealthy citizens are hiding money and is moving to seize it; separately, microchip smugglers are being caught importing banned chips into China.5:19The same restricted materials are also flowing to Russia, exposing a double standard that undermines the credibility of enforcement efforts.10:06China directly subsidizes the chemical companies supplying fentanyl precursors, structurally contradicting any meaningful cooperation with US enforcement.19
0:00Bank of America's report signals credit and debit card spending is surging to alarming levels, drawing parallels to pre-crisis excess.7:31Reenture's forecasts exclude recession scenarios entirely — all projections assume continued economic expansion with no downturn factored in.14:12The typical US home costs $535K against $30K annual rent, a price-to-rent ratio of 17.6 — roughly double the comparable figure elsewhere, signaling deep overvaluation.
0:00Bank of America's report signals credit and debit card spending is surging to alarming levels, drawing parallels to pre-crisis excess.7:31Reenture's forecasts exclude recession scenarios entirely — all projections assume continued economic expansion with no downturn factored in.14:12The typical US home costs $535K against $30K annual rent, a price-to-rent ratio of 17.6 — roughly double the comparable figure elsewhere, signaling deep overvaluation.20
0:00CGsociety's story begins with the CG Talk forums Leard Tio founded in Australia in 2001, a strong origin that would take over 20 years to unravel.4:25A successor effort attracted 5,000 users on its first day, yet its founders explicitly rejected the goal of rebuilding the original CGTalk model for a new era.8:21By this point CGsociety had become a nostalgic relic of the early CG internet, no longer the living industry hub it once was.
0:00CGsociety's story begins with the CG Talk forums Leard Tio founded in Australia in 2001, a strong origin that would take over 20 years to unravel.4:25A successor effort attracted 5,000 users on its first day, yet its founders explicitly rejected the goal of rebuilding the original CGTalk model for a new era.8:21By this point CGsociety had become a nostalgic relic of the early CG internet, no longer the living industry hub it once was.21
0:12Data and environment curation is the primary lever for improving LLM quality post-pretraining, the focus of Bespoke Labs co-founder Mahesh Sathiamoorthy.8:39The Open Thoughts curation pipeline begins with a large pool of candidate data and passes it through a defined sequence of filtering and curation stages.16:06Bespoke Labs' Curator tool is purpose-built for curating reasoning data, letting users programmatically define and run curation pipelines for post-training.
0:12Data and environment curation is the primary lever for improving LLM quality post-pretraining, the focus of Bespoke Labs co-founder Mahesh Sathiamoorthy.8:39The Open Thoughts curation pipeline begins with a large pool of candidate data and passes it through a defined sequence of filtering and curation stages.16:06Bespoke Labs' Curator tool is purpose-built for curating reasoning data, letting users programmatically define and run curation pipelines for post-training.22
0:05After covering attention the previous week, this lecture shifts to reinforcement learning and policy gradient methods.33:00Modern neural architectures combine attention and MLPs — these two components together form the backbone of current models.1:02:13Each major AI company maintains its own proprietary model, making cross-organization deployment significantly more complex than a shared-model world would be.
0:05After covering attention the previous week, this lecture shifts to reinforcement learning and policy gradient methods.33:00Modern neural architectures combine attention and MLPs — these two components together form the backbone of current models.1:02:13Each major AI company maintains its own proprietary model, making cross-organization deployment significantly more complex than a shared-model world would be.23
0:01Agents broadly fail to support async MCP tasks; the session opens the question of why that adoption gap persists and what it costs agent workflows.10:38The live dashboard shows all running processes tracked in real time, demonstrating that the system maintains full state visibility across concurrent async operations.20:02With 13 seconds remaining the demo is still live; a planned V2 switch is skipped, but the high-level architecture it would show holds across both versions.
0:01Agents broadly fail to support async MCP tasks; the session opens the question of why that adoption gap persists and what it costs agent workflows.10:38The live dashboard shows all running processes tracked in real time, demonstrating that the system maintains full state visibility across concurrent async operations.20:02With 13 seconds remaining the demo is still live; a planned V2 switch is skipped, but the high-level architecture it would show holds across both versions.24
0:05Three to four upcoming lectures will cover large language models.34:54Raising temperature softens the output distribution, shifting mass to the long tail and enabling more diverse token generation.1:05:49Single-head attention is just the baseline; multi-head attention adds further complexity needed to complete the full mechanism.
0:05Three to four upcoming lectures will cover large language models.34:54Raising temperature softens the output distribution, shifting mass to the long tail and enabling more diverse token generation.1:05:49Single-head attention is just the baseline; multi-head attention adds further complexity needed to complete the full mechanism.25
0:05Lecture agenda: unsupervised learning via GMM/EM and PCA, then reinforcement learning—PPO's clipped objective and framing language models as MDPs.36:26PPO's clipped objective caps policy updates to a trust region; the advantage function measures how much better an action is than the current policy's baseline.1:03:58Language models map onto MDPs: each token is an action, the context window is the state, enabling reinforcement learning directly on generation.
0:05Lecture agenda: unsupervised learning via GMM/EM and PCA, then reinforcement learning—PPO's clipped objective and framing language models as MDPs.36:26PPO's clipped objective caps policy updates to a trust region; the advantage function measures how much better an action is than the current policy's baseline.1:03:58Language models map onto MDPs: each token is an action, the context window is the state, enabling reinforcement learning directly on generation.26
0:10The speaker is at the seminar for the third time, still associating it with its founder Terry Winograd.25:24Reading fluency evidence suggests the text approach under discussion is probably not a good idea.47:50The juxtaposition frame — asking what text-and-visualization combinations actually do — is the productive lens for analyzing them.
0:10The speaker is at the seminar for the third time, still associating it with its founder Terry Winograd.25:24Reading fluency evidence suggests the text approach under discussion is probably not a good idea.47:50The juxtaposition frame — asking what text-and-visualization combinations actually do — is the productive lens for analyzing them.27
0:00Career-shaping advice: do what makes you happy; if it doesn't, do something else. Those words stayed with Tatum.19:48The NBA treats franchise relocations with extreme sensitivity because moving a team between cities carries serious consequences.37:26VR via Oculus delivers compelling game experiences but requires full headset immersion; the future of VR for NBA viewing remains uncertain.
0:00Career-shaping advice: do what makes you happy; if it doesn't, do something else. Those words stayed with Tatum.19:48The NBA treats franchise relocations with extreme sensitivity because moving a team between cities carries serious consequences.37:26VR via Oculus delivers compelling game experiences but requires full headset immersion; the future of VR for NBA viewing remains uncertain.28
0:05The course shifts to reinforcement learning this week, with today's session covering foundational RL basics.34:25In RL, a policy is abstractly a mapping from state S to action A.1:04:53The policy gradient derivation writes expected value as an integral of probability times value, then moves the gradient inside the integral.
0:05The course shifts to reinforcement learning this week, with today's session covering foundational RL basics.34:25In RL, a policy is abstractly a mapping from state S to action A.1:04:53The policy gradient derivation writes expected value as an integral of probability times value, then moves the gradient inside the integral.From your X · originals only
4
HANDLES
POSTED
POSTED
8
ORIGINAL
POSTS
POSTS
@BrianFeroldi ×3@JesseOlson ×3@gregisenberg ×1@Investanswers ×1
36 ITEMS FROM 23 OF 806 TRACKED CHANNELS — EVERY POST READ & RANKED