Briefing

DRAM supply dynamics in 2026 are being reshaped by AI demand: three producers (Samsung, SK Hynix, Micron) share a finite wafer pool and HBM—used in AI servers—consumes roughly 3–4× the wafer capacity per GB of standard DDR5. That imbalance has pushed HBM to ~30% of DRAM revenue while only supplying ~8% of bits, contributed to consumer/laptop RAM doubling in price this year, and left SK Hynix with essentially no sellable HBM capacity through 2026. Micron has reallocated wafers (walking away from Crucial) and locked multi‑year, prepaid contracts covering ~50% of revenue at floor prices above prior cycle margins, which the author argues breaks the historical commodity cycle. With HBM4 increasing stacking (12→16 dies) and buyers fronting cash, memory looks more like a constrained toll road for AI capex than a tradable commodity, a thesis the market may not yet have fully priced.

NVIDIA DGX Spark is presented as a practical, low‑cost (listed $4,699) way to run models that exceed conventional GPU VRAM limits: the unit uses a GB10 (Grace Blackwell) SoC with 128 GB unified memory, ConnectX‑7 200GbE QSFP, 4 TB encrypted NVMe and ~1 PFLOPS FP4. The author and community testers (@MiaAI_lab) recommend a 2x Spark cluster as the "sweet spot": direct 200GbE QSFP link, static IP netplan examples, passwordless SSH, NCCL verification and matching drivers. Benchmarks show Qwen 3.6 35B NVFP4 at 256k context ≈110 tok/s (Apr 2026) on 1x Spark, DeepSeek V4 Flash DSpark hitting 1M context and 40–45 tok/s (≈79 tok/s with 3 concurrent sessions) on 2x, and GLM 5.2 NVFP4 ~20 tok/s at 1M context on 4x. The post includes vLLM Docker commands for single and distributed serving and operational guidance (--gpu-memory-utilization 0.85, --max-num-seqs 4, cache safetensors, warm up, pin digests) plus notes on scaling to 4–8 Sparks with 200GbE switches (e.g., MikroTik CRS504).

TSMC and Winbond will jointly develop Wafer-on-Wafer (WoW) 3D stacking: Winbond supplying DRAM wafers and TSMC stacking them with logic wafers for AI chips. The move, reported 2026-06-28, is positioned as TSMC’s strategy to diversify beyond Samsung, SK Hynix, and Micron and build a Taiwanese memory supply chain while enabling Winbond’s push into AI server and HPC markets.

Oracle became the first S&P 500 company to explicitly blame AI in its SEC filing: cutting 21,000 roles (13%) and taking a $1.8B charge after pilots that replaced 47 DBAs with 3 senior architects plus AI (catching ~94% of issues and shrinking 6‑week tasks to 6 hours). Oracle says savings will fund a $50B FY2026 AI infrastructure buildout; the post frames this as a replicable Fortune 500 playbook amid ~90,000 tech layoffs by March 2026, >70% citing AI.

DeepSeek released DSpark (and the DeepSpec training/eval stack) on 2026-06-28, open-sourcing the code and checkpoints under an MIT license and claiming production deployment across millions of users. DSpark addresses the token-by-token latency of large LLMs by combining a parallel drafter backbone with a lightweight Markov head so drafts are both fast and token-coherent, yielding 26–31% better accepted-token lengths versus Eagle3 (Qwen3 families) and 16–18% versus DFlash. It also adds a confidence-prediction head that estimates rejection probability and dynamically trims verification length based on server load to avoid wasted GPU verification. DeepSeek reports per-user generation speed gains of 60–85% on V4 Flash and 57–78% on V4 Pro (vs MTP-1), with aggregate throughput improvements of 51–400% under high concurrency. The paper lists founder Liang Wenfeng and Peking University collaborators; checkpoints are on Hugging Face and training code on GitHub, and the framework already supports Qwen3 and Gemma models.

High‑purity CO2 procurement for semiconductor supercritical cleaning is under strain after June 26, 2026 reports showed inventories falling below the usual combined one‑month buffer. Reduced CO2 feedstock from lower oil refinery and petrochemical utilization—linked to US/Iran/Middle East crude instability—has driven ~20% price rises and constrained supplies that cannot be quickly ramped up, risking continued shortages through year‑end.

La inteligencia artificial está disparando la demanda de energía y obliga a empresas como xAI y OpenAI a planear centros de datos de cientos de megavatios para entrenar modelos; un clúster puede consumir tanta potencia como una ciudad de 100,000 habitantes. Las compañías están cerrando acuerdos privados con productores de GNL y reactores nucleares modulares, generando capacidad y empleos en Texas y Nevada, según el autor.

The Kobeissi Letter reports the Trump Administration on June 27, 2026, reached a deal with Anthropic to permit release of Mythos 5 to roughly 100 companies and federal agencies (per CNBC). Anthropic had disabled Fable 5 and Mythos 5 earlier that month for export-control compliance; senior Anthropic staff flew to Washington after a two-week standoff, and the post predicts industry-wide effects.

Jukan (@jukan05) cites Gavin S. Baker’s All‑In comment at 1:04:06 that "DRAM is the most important bottleneck" and that Elon is focusing the Terafab on memory, not lasers, capacitors, power semis, NAND or HDDs. Jukan says this validates DRAM’s strategic value, rejects the idea it’s easily replaced by Chinese products, and argues Micron (MU) could exceed Nvidia in importance because inference ROI depends on memory, not GPUs.

Shruti (@heyshrutimishra) reports humanoid production has reached 15,000, with two years to build the first 1,000 and only three months for the most recent 5,000, signaling a steep adoption curve. At Longcheer in Nanchang, AgiBot's G2 ran an 8‑hour shift, self-corrected a conveyor deviation, ran 10 hours total, and sorted >3,000 units at 100% success while interfacing with factory software; she warns the U.S. is underrating the pace.

España dispone de recursos solares y eólicos entre los mejores de Europa pero sufre algunas de las tarifas eléctricas más altas, según @martinvars (28 junio 2026). Culpa a decisiones de política —cierres programados de centrales nucleares operativas, impuestos y peajes elevados y una red dependiente del clima—; el apagón reciente lo evidenció. Afirma que la energía cara es resultado de una decisión de gobierno y puede cambiarse.

Micron (MU) is positioned as potentially more critical than Nvidia for AI inference: @jukan05 (2026-06-28) argues inference ROI hinges on memory capacity, not GPU count, because memory bottlenecks leave GPUs idle. A reply from @TicTocTick predicts MU falling from "now 1200" to 700 and repeats the slogan "RAM is NOT GPU."

Aluminum-capacitor makers in China and Japan have launched a coordinated round of price increases as H2 2026 begins. Rubycon — Japan's third-largest maker — announced a price rise effective August 1, 2026 for electrolytic, solid-aluminum and film capacitors and warned of a possible second hike, citing sustained high crude oil, surging metals (aluminum, copper, tin) and supplier-wide cost increases across materials and logistics. In late June 2026 Jiang Hai, China’s leading aluminum-capacitor supplier, issued its own notice raising prices for aluminum capacitors, film capacitors and supercapacitors because aluminum foil, chemical inputs, carbon powder and electricity costs have exceeded prior pricing models. The post highlights that Nippon Chemicon, Nichicon and Rubycon now all are raising prices, that these makers ship substantial volumes into AI server supply chains, and that the combined upstream cost pressure makes it easier for Taiwanese makers to pass costs on to customers.

Gary Marcus (2026-06-28) warns AI will be low-margin and costly—“more like the airline industry”—and calls pouring “trillions” into it likely unwise. He amplifies Xiaoyin Qu’s scenario that China could undercut OpenAI and Anthropic by giving away top-tier models and exporting cheap compute (cheap electricity plus Huawei closing the chip gap), driving inference costs from “hundreds of billions” toward “almost zero.”

The author recounts meeting a friend who runs an AI SaaS generating $70k MRR but only ~30% profit (~$21k/month) after infrastructure, AI-agent, and Google/Meta ad costs. He contrasts that with his higher-margin B2C apps at lower MRR and frames the strategic choice: prioritize immediate high-margin cashflow or build a more sellable SaaS asset.

John Loeber (2026-06-28) argues that all future software must be formally verifiable to make cybersecurity possible and to support codebases 100x+ larger as AI increases scale. Jesse Michael Han quotes Gwern saying 2026-era LLMs give Lean worse baseline loss but superior scaling exponents, so Lean rewrites could eventually deliver major correctness and security gains.

Brian Armstrong outlines operational tactics to keep AI spend flat while token usage grows: default engineers to open-weight models (GLM 5.2, Kimi 2.7), preprocess and route prompts to the right model, and implement cache-aware requests (LibreChat hit rate 5%→60%). These changes reportedly cut AI spend nearly in half without suppressing usage.

El tuit de @martinvars (28-06-2026) avisa que España tendrá otra vez más muertes que nacimientos y que la fuerza laboral que sostiene a los jubilados se reduce. Rechaza soluciones como inmigración o cheques por hijo y exige atacar raíces: vivienda asequible, empleos tempranos y mejor pagados; su receta política es construir mucho más, desregular y reducir cargas para que los jóvenes puedan formar una familia.

@bcherny maps five product-role archetypes he sees on the Claude Code team—Prototyper, Builder, Sweeper, Grower, Maintainer—and argues people often span 2–3 archetypes. He notes these archetypes cross engineering, design, PM and DS functions and recommends specific mixes (1+2+3, 2+3+4+some5, 3+4+5+some2) depending on product stage.

Anton Osika highlights that his friend Kristian Rønn (@Kristian_Ronn) is developing cryptographic verification to ensure superintelligence isn't misused, letting both open-source projects and commercial labs prove compliance without revealing secrets or private data. Osika casts the issue as a verification shortfall—'no one can verify what anyone else is doing'—and links to Billy Perrigo’s TIME piece (June 23, 2026).

Black Spark claims it infiltrated the Alabuga SEZ in Tatarstan, stole personnel databases tied to Geran-2 and Geran-3 assembly, hacked the site (archived 27 June 2026) and secretly embedded "surprises" in recent UAV batches that will produce "unpleasant consequences" on launch; Alabuga allegedly shut servers and FSB agents are now checking staff, per Michael D. Weiss.

Levie (June 28, 2026) warns that imminently open 'mythos level' cyber-security models will spawn alternative tech stacks and shift economic/control leverage away from the US; he says withholding models won’t make the US safer and instead risks asymmetric disadvantage, and points to a Polymarket claim that Zhipu AI matches Claude Mythos at bug-finding as proof China can catch up.

Gary Marcus highlights Ford's failed experiment replacing engineers with AI: after cutting more than 5,000 jobs since 2020 and relying on AI-adjusted designs, Ford rehired or promoted 350 experienced engineers when quality collapsed. The company now tops JD Power for the first time in nearly 20 years, plans 100,000 more AI tests, yet CEO Jim Farley still predicts mass white-collar displacement.

SurmountInvest recounts the June 23, 2026 KOSPI rout — a 9.99% drop to 8,203.84 after Samsung (-12.3%) and SK Hynix (-12.5%) triggered a circuit breaker — and warns that record retail net buying (11.55 trillion won) funded by a record 38.48 trillion won in margin debt and new 2x ETFs leaves retail overleveraged and at systemic risk.

Brad, the CEO of O'Reilly Automotive, started as a 17-year-old parts specialist in Wagoner, OK in July 1996 and rose through every store role before age 20. Being passed over for district manager at 23 taught him to develop others; he became CEO in 2024 as O'Reilly scaled massively and institutionalized store-to-executive promotion.

@cramforce (2026-06-28) says a brief window exists where defensive measures—sharing source code and using automated harnesses—outperform black-box pen testing. He urges teams to run deepsec (vercel-labs/deepsec) with frontier coding models immediately to harden applications. Guillermo Rauch warns Mythos/Sol-style tools can be weaponized by adversaries, threatening US firms.

Author @jvisserlabs argues that June 2026 pain among the 'Mag 7' hyperscalers reflects rotation, not collapse: AI is consolidating as capital flows to healthcare, small caps, equal-weight, AI infrastructure receivers, and memory (Micron) amid shortages. He warns token usage could surge via loops/tags and predicts a shift from infrastructure to AI applications.

Chubby♨️ argues Anthropic’s CEO lobbied for AI regulation as a strategic move to shift blame away from the company if 'Mythos 5' later causes damage in China. The post alleges the CEO sought regulations 'few months ahead of an IPO' to enable blaming the US government rather than Anthropic.

Josh Rauh warns that California’s Property Seizure Act — labeled the 'Billionaire Tax' and highlighted by Chamath Palihapitiya and Hoover Institution media — would require a 5% annual cash levy on all tangible and intangible property, trigger decade-long litigation likely to reach the Supreme Court, spur business flight, deepen the budget hole and risk pension resets, higher taxes and near-bankruptcy.

@aymanalabdul lays out a five-level AI maturity model: from AI as a thought partner (Level 1) to agents managing agents with compounding feedback loops (Level 5). He argues most CEOs remain at Levels 1–2 yet label themselves 'AI native', and that true competitive separation will come from firms advancing toward Level 5 autonomy.

Kimmonismus (2026-06-28) challenges claims that “a new model from Zhipu AI” matches Fable5 on cybersecurity, finding only a Wall Street Journal piece about GLM-5.2. WSJ says GLM-5.2 can match top U.S. models in some bug-finding tasks, and 360 Security claims Tulongfeng rivals Anthropic’s Mythos, suggesting either underestimated GLM-5.2 or misleading reporting.

Argentina closed its first sustained fiscal surplus in more than a decade, the post asserts, after measures including big spending cuts, elimination of entire ministries, and price liberalization. The author reports monthly inflation dropped to single digits without bailouts or accounting tricks and argues Javier Milei demonstrated that fiscal discipline—spending less than revenue—produces prosperity, a lesson for Spain and the U.S.

Author @martinvars recalls getting lost at age 5 in Villa Gesell, Argentina, where a tall stranger hoisted him onto his shoulders and a clapping crowd quickly reunited him with his parents; he contrasts that with life in the USA (2026-06-28 post), asserting Americans wouldn’t intervene similarly and that US society lacks such informal safety mechanisms.

Martín Vars sostiene que la IA médica debe orientarse a ampliar el acceso: al ayudar a leer imágenes, priorizar casos y detectar riesgos tempranos, estas herramientas multiplican el alcance del clínico y cubren "territorios donde hoy no hay ninguna" atención. En Certuma ven diariamente cómo esa tecnología llega a pacientes que viven horas de distancia y esperan meses por consulta.

A spinning clear-glass disc mounted before a camera lens—rotated to about 3,000 RPM (50 rev/s)—uses centrifugal force (~500 g at the edge) and shear to atomize raindrops so images stay clear, while rotation can create a thin vacuum that lifts snow. The hard part is balancing the disc to near-zero wobble; the concept was patented in 1917 by two British engineers and is still used in modern broadcasts.

El fundador y autor @martinvars (publicado 2026-06-28) sostiene que, tras 30 años creando empresas, nunca vio una herramienta que premie tanto al que ya sabe operar: la IA permite rediseñar procesos, eliminar pasos, fusionar traspasos y detectar errores antes de costo. Aplicaciones superficiales tipo chatbot cambian poco; la ola favorecerá a operadores serios sobre modas.

@deanwball on 2026-06-28 argued that First Amendment litigation will decide whether creating, distributing, and using frontier AI (large language models) is protected speech; he called for concrete fact patterns to establish standing for non-lab parties, to abandon simplistic 'code is speech' approaches, and to marshal top legal talent despite national-security tensions.

@deanwball argues (2026-06-28) that while the First Amendment (1A) will constrain state intervention and could complicate finely detailed alignment rules, it does not make AI regulation impossible. He urges regulators to target frontier labs as entities rather than models, framing this limit as an acceptable tradeoff to preserve American popular sovereignty and 1A protections.

Austria's outreach to Anthropic (reported 2026-06-28) offers legal certainty, market access and capital to host the company in the EU, but the author argues this is a desperate move: U.S. regulation and control of top compute (and possible designation of Fable and Mythos as national-security assets) would still keep Europe cut off, so Anthropic will stay near U.S. compute.

Dean W. Ball argues (2026-06-28) that the central legal battles over AI will be First Amendment cases: courts must determine whether the creation, distribution, and use of frontier AI are protected expression and who outside labs has standing to sue. He warns against relying on 'code is speech' rhetoric, notes national-security complications, and urges top legal minds to address these questions given that the underlying tech is a large language model.

The African elephant’s ~2,000 olfactory receptor genes give it far superior smell: it can detect water 12 miles away, locate a ripe tree from kilometers, and differentiate human groups by scent. In the posted video she constantly resamples the air with her trunk, had already mapped the fruit buffet, and entered backwards to let people think she was surprised.

Laurel's spreadsheet (described by Aakash Gupta on 2026-06-28) converts one person's tacit AI workflow into a 200+ row Company OS: each task has a one-line description plus an explicit Negative Example that draws boundaries (Product split into five subfunctions). That labeled structure lets agents and new hires route Slack feature requests and enforces AI-first norms (Agentic: manual specs are off‑spec).

@ivanfioravanti (2026-06-28) previews experiments with Qwen3.6-27B-MTP-GGUF on four 64×64 images (three plain, one MTP), claims “An image is worth 16x16 words,” and says DGX Spark excels at prefill but is too slow at decode. They emphasize real-life use sends smaller chunks cached on SSD and are working to enable flawless 128K context on M5 Max and DGX Spark with llamacpp.

Aakash Gupta argues that Figma — the defining product-design tool since 2016 — has been demoted to step 4 in a new five-step workflow (LLM → user call → AI prototype → Figma → live-iteration) popularized by Ramp. He claims generative tooling (2026) removes the craft bottleneck, shifting value to taste, systems thinking, and judgment; Codex-powered PRs exemplify this upstream shift.

Lulumeservey posted a short reaction video on 2026-06-28 agreeing with Scott Wu's claim that AI will so change the nature of work that future generations won't recognize today's office jobs as 'work.' The post highlights Scott's contrast between 'pushing buttons' and meetings versus ancestral manual labor like farming and handmade clothing.

@ivanfioravanti expresses a desire to work at SpaceX to access and experiment with the latest models and hardware. They quote Elon Musk's June 28, 2026 tweet announcing Grok 4.5—based on a 1.5T V9 foundation model with Cursor data, in private beta at SpaceX and Tesla, with early evaluations near or exceeding Opus, ongoing RL gains, and monthly from-scratch releases in 2026.

Dan McAteer predicts GPT-5.6 will ship and Fable will return the week of June 28, 2026, citing Axios that export controls on Fable could be lifted immediately and won’t favor Anthropic over OpenAI. He argues Fable is most capable but slow and costly, while GPT-5.6 Sol is nearly as strong, 40% cheaper per output token, and may have aided solving the Erdős problem.

Austria's bid to make Anthropic 'European' argues Europe has some existing AI compute and will expand, but the decisive barrier is likely US government consent. The author proposes an EU-controlled Anthropic subsidiary as a realistic compromise—granting EU jurisdiction and reducing Chinese influence—yet predicts Anthropic will probably stay in the US where compute and supply are concentrated.

OpenAI's GPT-5.6 preview (announced 2026-06-28) unveils three tiers—Sol (frontier), Terra, and Luna—but at the U.S. government's request none will go straight to the public; Sol will launch as a limited preview for approved partners and is billed as the company's most capable model for coding, science, and cybersecurity. The post links this to Anthropic's Fable 5 (pulled after three days) and Mythos (never public), arguing frontier models are increasingly invite-only and asking whether government-gated access is now the norm and if Sol will surpass Fable's brief appearance.

Corey Ganim outlines five Second Brain-as-a-Service products (law, tax, B2B sales, restaurant franchises, tutoring), a 45-minute build recipe (raw/wiki/outputs; schema file; AI-built wiki), and a GTM/pricing plan targeting five customer types with starter builds at $1.5k–$3k, advanced builds $5k+, and $300–$1k/mo maintenance; he cites @TheViableEdge commercializing the model.

Chamath (@chamath) warns California’s Property Seizure Act—marketed as “The Billionaire Tax”—is on the ballot and would, he claims, require a 5% annual cash seizure of all tangible and intangible property. He predicts decade-long litigation likely to fail in the Supreme Court, an exodus of business builders, large revenue losses, deeper deficits, higher taxes on middle-class Californians, and possible near-bankruptcy and pension resets.

Shruti (@heyshrutimishra) reports her keyboard usage has fallen to 20% as she dictates almost everything and claims a 5× speedup. She highlights FluidVoice, a free, local macOS voice-to-text tool (GPLv3) with no cloud or accounts, arguing open source plus local models removes trust issues and makes the keyboard "officially dead."

Jason C. (@jasonc_nc) warns that a proposed housing approach would weaponize state power to displace owners, subsidize chosen new owners and run buildings below cost, eroding property-tax revenue, stopping new construction and increasing subsidy demand. He recommends letting tenant associations form limited-equity co-ops from foreclosed buildings (Housing Development Fund Corporation model) and urges fully funding transit.

@bscholl argues he is pro‑business but warns some firms commit real harm by weaponizing government to block rivals. He claims Boeing drove 1970s opposition to supersonic travel after its jet cancellation and contends today’s AI “doomerism” is the same tactic—incumbents using red tape to stop fast innovation—and must be called out and outlawed.

Jason C (@jasonc_nc) argues that virtually every great neighborhood predates the American Planning Association’s predecessors and that increasing planning control has steadily worsened urban quality (post dated 2026-06-28). Carol Walsh (@CarolWalshReal1) offers Esplanade Street as a counterexample, detailing one-way side streets, a bike lane, landscaping, set-back mixed uses, trees, and high walkability.

GLM-5.2, released by Chinese developers and highlighted on 2026-06-28, is presented as matching OpenAI and Anthropic models and is called the “second DeepSeek moment.” Quoting Innovation Council and David Sacks, the post argues this parity creates an urgent, competitive AI race and warns against delaying U.S. model deployment or hampering companies’ progress.

@levelsio reports coding almost entirely on his VPS with Claude Code for nearly a year, praising instant live edits, device switching, overnight runs via /goal, and rapid go-live from scratch. He experienced only two brief outages (~10 seconds) in 12 months, previously reduced deploys from ~1 minute to ~3 seconds, and argues AI coding agents will run in the cloud, while advising backups and staging for teams.

Jason (@jasonc_nc) argues fire departments should embrace being "emergency services departments" with only a "tiny" portion devoted to fire; by making minor equipment changes they could directly improve public safety and use urban streets more effectively. He specifically calls out the LAFD and says an overemphasis on response speed, treated as the sole safety metric, leads departments and unions to gaslight the public and fabricate impacts.

Sarvesh Shrivastava (@bloggersarvesh) argues that Claude Cowork can replace a $10k SEO team for $20/month by shifting from vague chatbot prompts to a structured SEO system: preload full business context, ask strategist-level questions, run competitor gap analysis, cluster keywords by intent, and optimize for AI search. He published 20 free prompts and an infographic on 2026-06-28.

levelsio published a dynamic chart that constantly adds Bitcoin 4-year cycles, marking October 31, 2026 as the next cycle start and theoretical bottom. He lists crash years 2014, 2018, 2022 and 2026, projects an up-cycle from 2027–2031, and links the pattern to the quadrennial halvening and repeated sentiment-driven selloffs.

Andrew Ambrosino, who leads the Codex desktop app at OpenAI, reports the product grew 6x since February to over 5M weekly active users and is used by nearly all OpenAI employees. He frames his mission as building "the best desktop app" and argues timing and model upgrades (November vs. February) determined the app's success; the interview covers PM "zone defense," design limits of AI, "taste," workflows, and Codex+ChatGPT plans.

Chubby♨️ (@kimmonismus) argues that the U.S. embargo on Fable 5 and GPT-5.6 was not the result of Dario Amodei’s so‑called “fearmongering” but rather a calculated national‑security decision: while Anthropic’s February refusal to cooperate with the U.S. Department of Defense and other communication failures are fair criticisms, the U.S. government—backed by advisers, NSA and other experts—would not halt a technology and impose massive financial and investment requirements solely because of a CEO’s statements. The author believes regulators fear the models could be distilled or otherwise transferred to China and used for cyberattacks, espionage, or large‑scale harm; they say the resulting mitigation demands are enormous (compared rhetorically to the Manhattan Project). Chubby♨️ expresses concern that public access may be curtailed and endorses open source as the preferable response.

FreeLLMAPI aggregates 11 free LLM providers into one OpenAI-compatible /v1 endpoint and claims roughly 1.7 billion free tokens per month. It provides an automatic fallover router, single API key, encrypted key storage, per-key usage tracking, and support for chat, embeddings, images, audio, tools, and streaming to reduce inference costs for AI side projects.

Nan Ransohoff announced Intercept on 2026-06-24, a $500M philanthropic effort to eliminate respiratory infections by funding broad-spectrum preventatives and air-cleaning technologies. She highlights a global burden—5% of life sick, ~1M deaths/year, $600B lost—and says a convening of ~40 experts found technical complexity and sustained funding gaps are the main obstacles.

Michael Chomsky's post celebrates 1,000 listeners for Pod 3 of the Directionally Wrong podcast (published 2026-06-26), thanks guest @densumesh, and announces upcoming distribution on YouTube, Spotify and other platforms. Changelog notes audio quality improvements and removal of 'microphone condoms.' Co-promotion by Nick Khami lists episode topics — Dan Abramov joining Vercel, Chinese AI, MCP failure, corgi/papermark, and code-review startups — and links to the live broadcast.

Michael Chomsky (@michael_chomsky) argues that Sakana AI's creation of a 'model router'—called the future of Japanese AI—typifies a 'Money Scaling Intelligence Company' approach: more compute (money) produces more intelligence. He advocates model routing (oracle LMs aiding small models), background compute/dreaming, and vertical/horizontal scaling while rejecting the idea of models as fixed artistic works like Fable.

@mhdempsey (2026-06-23) says recruiting choices now tilt decisively toward relative positioning, earnings and time-to-liquidity. He contrasts younger cohorts who obsess over sequencing (logo, equity, title) with mid/experienced hires driven by relativity and prior-bubble scar tissue. These behavioral shifts are altering hiring incentives and time horizons, risking short-term optimizations that may become harmful long-term.

@mhdempsey wrote a brief tribute on 2026-06-25 describing how he 'grew up in tech' reading @om's blog and admired the photography Om and @bijan shared—often documenting walks across Boston, San Francisco, and New York City. After finally meeting Om, they bonded over investing, photography, and shared heart‑condition experiences; he concluded, 'Rest in peace, Om.'

Akio Morita refused a 1955 Bulova offer to buy 100,000 TR-55 radios if Sony's name was removed. With Sony only months old, Morita overruled partners who wanted the revenue, predicting Sony would be as famous as Bulova in fifty years; he later called it his best decision, while Bulova was sold to Citizen for $250 million in 2008.

Elon Musk is presented as a 55-year-old ‘guardian Angel’ whose ventures—SpaceX, Tesla, Starlink, Neuralink, Optimus, and AI/Grok—are cast as deliberate, often confrontational efforts to secure humanity’s future. The post credits SpaceX with reusable rocket landings and a multiplanetary mission to preserve consciousness; Tesla with forcing an industry-wide shift to sustainable energy via open-sourced patents; Starlink with restoring connectivity to Ukraine; Neuralink with enabling a paralyzed man to move a cursor by thought; and Optimus with targeting scarcity by automating backbreaking work. It asserts Musk paid $44 billion to buy Twitter to expose censorship and restore a public square, created Grok as a truth-seeking AI, backed Trump at personal cost, exposed institutional coverups of child abuse, called out unelected judicial power, warned about declining birth rates, and emphasizes that his motivation has always been humanity rather than money.

swyx framed the AIEWF "Stress Curve" as the Gini coefficient over time from ticket launch to sold out; AI Engineer replied that the 2026 World's Fair is completely sold out — leadership track, workshops, late-bird tickets — while offering limited overflow tickets, 65 free SF side events, and a Moscone orientation 5–9pm.

PolyHelper positions itself as the 'pro version' of Polymarket, adding features Polymarket won't implement — copytrading, trader tracking, and advanced analytics — while letting users keep the Polymarket interface. @0xd1namit (2026-06-28) argues this is necessary because Polymarket targets retail and cannot overload its native UI.

@aakashgupta explains that Liebig's 1835 ammoniated silver‑nitrate + reducing‑sugar method plates silver atom-by-atom across glass, producing a continuous film about 100 nanometers thick—thinner than a wavelength of red light. The sugar's aldehyde reduces Ag⁺ to Ag⁰ while oxidizing to gluconic acid, replacing earlier mercury‑amalgam mirror techniques that caused toxic exposure.

The 2026 World's Fair is completely sold out and promoted as "the largest AI industry expo on earth," with the Leadership track for CTOs & VP AI's, workshops, and all late-bird tickets gone. Organizers opened limited overflow expo/engineering tickets (no guaranteed seating), note 65 free side events around SF, and host New Engineer Orientation at Moscone tonight (June 28, 5–9pm).

Yishan (@yishan) used the authors' publicly released evaluation framework (github.com/aiden-ygu/health-...) and posted his reproduction at github.com/ywong137/health-a..., re-running the paper's radiology-interpretation tests on both the original models and current frontier models. He found little baseline drift and measured an improvement from 69/100 (prior best) to 79/100 for GPT-5.5 Pro, but stresses this numeric gain still falls short of the paper's stricter clinical-readiness requirements—robustness under perturbation, proper abstention when data are insufficient, and clinically valid reasoning without hallucinations—tests he could not replicate qualitatively. He notes some testing data remained closed due to JAMA/NEJM copyrights, praises the authors for releasing what they could, and argues that open-sourcing evaluation code is essential so others can reproducibly run benchmarks on newer models over time.

Gary Marcus posted a single clown emoji ("🤡") in response to Acyn's June 28, 2026 tweet reporting that Fox News Media apologized after Kevin O'Leary retracted claims about opponents of his planned Utah data center. O'Leary said he has "no evidence" the opponents were funded by or coordinated with the Chinese Communist Party; Fox acknowledged the error.

cryptopunk7213 and Brian Armstrong argue the near-term value lies in model selection per task, multi-model routing/harnesses, caching, and AI-cloud memory/DB management as aggregation becomes the bottleneck. Armstrong reports defaulting to GLM 5.2 and Kimi 2.7, 91% of employees never hit caps, a LibreChat cache hit improvement from 5%→60%, and nearly halved AI spend while tokens grow.

Daniel Mac (@daniel_mac8) predicts GPT-5.6 will ship and Fable will return the week of 2026-06-28, citing Axios reporting that export controls on Fable may be lifted and the government won’t favor Anthropic over OpenAI. He frames Fable as the most capable but slow/expensive model, GPT-5.6 as nearly as capable and 40% cheaper per output token, and highlights combining Fable (coding) with GPT-5.6 (agentic experiments); tester Mikhail Parakhin corroborates 5.6 beating Opus 4.8 and being better at agentic tasks while Fable is stronger on coding.

UnveiledChina and @dennisw5 say the ODNI is drafting a public, unclassified report — mandated by the FY2026 Intelligence Authorization Act and due before December 2026 — that will expose hidden fortunes of Xi Jinping and the 25-member Politburo; the thread cites CRS data linking over $700 million to Xi relatives and argues revelations about Xi’s sister and offshore holdings will undercut CCP propaganda and public trust.

Jason (@jasonc_nc) contrasts emergency response metrics (Paris: 12 min fire, 7 min EMS; Houston 90th‑percentile: 13 min fire, 15 min EMS) with traffic‑safety opportunities: US lane standards (12–14 ft lanes, 26 ft two‑lane minimum) hinder smaller apparatus and calming measures, while narrower European lanes (Stockholm ~9.8 ft) and road diets cut crashes and could save far more lives than further shrinking fire response times.

Author @jukan05 cites IBM CEO Arvind Krishna’s podcast claim that enterprise AI incurs a net loss in year one (making 80% of firms look failed) but delivers a 10x return in year two. The post argues CSPs’ revenues will expand at that inflection, allowing absolute CAPEX growth with falling CAPEX/revenue ratios and a virtuous flywheel, so cash flow must account for future AI-driven gains.

Andrew Ambrosino, who leads development of OpenAI’s Codex desktop app, reports a 6x increase in usage since February to over 5 million weekly active users and near-universal internal adoption at OpenAI. The interview covers product timing (November vs. February), OpenAI’s "zone defense" PM approach, design limitations of AI, professional "taste," and plans for tighter Codex–ChatGPT integration.

Andrew Ambrosino leads OpenAI’s Codex desktop app and aims to build “the best desktop app that has ever existed.” Codex grew 6× since February 2026 to over 5 million weekly active users and is used by nearly all OpenAI employees. In a June 28 interview he covers PM “zone defense,” design limits of AI, timing effects from model updates, and Codex+ChatGPT plans.

The post by @kimmonismus (2026-06-28) claims modern AI was built on the open internet’s “billions of texts, forum posts, code repositories, Wikipedia articles, scientific papers, blogs, and discussions,” and insists AI must not be sequestered behind proprietary walls. They frame open source as a geopolitical power issue and, despite transparency critiques, remain strongly pro–open source.

Aakash Gupta argues Rob Gronkowski is correct: a $60,000 private jet across the U.S. is poor ROI because $60k buys 60 first‑class coast‑to‑coast tickets. A 5‑hour flight saves ~45 minutes—about $1,333 per minute—while modern lie‑flat suites, Michelin meals and private‑like service make commercial travel sufficient; private pays only for <2‑hour hops, groups of 6+, or complex connections.

Mixture of Agents (MoA) ensembles parallel LLM drafts and an aggregator synthesis; Together AI's MoA hit 65.1% on AlpacaEval 2.0 vs GPT-4o 57.5% using open models, but the exam wasn't released and was written by the lab. MoA recovers collective knowledge and favours judge-driven benchmarks that reward long answers, while adding sequential-layer latency for only single-digit accuracy gains.

Michael Dempsey (@mhdempsey) (2026-06-23) observes that many younger people in tech rush career decisions, measuring opportunity cost over 12–24 months and fearing they’ll miss out as peers "kill it." He quotes a boss about plans giving way to "survival," urges patience — craft a career over time — and insists he won’t rush his own "life’s work."

@maxspero calls out an account for preaching authenticity while, he claims, “most of the posts are AI-generated,” and quotes Hunter Biden’s June 24, 2026 thread about “NY yesterday.” Biden’s seven lessons stress measurable authenticity, skepticism of establishment endorsements, conviction over triangulation, cost-of-living as top issue, rejecting the middle, courting the base, and fighting within the party (Mamdani).

@ianmiles (28 June 2026) argues voters have opposed more migration for decades but successive governments continue to "flood" the UK with migrants. He calls for immediate mass deportations and remigration, claiming Britain’s identity is at stake. The post contrasts with a Home Office announcement of limited, screened safe/legal refugee and sponsorship routes.

Author @meishato debunks the recurring claim that '9 million women were burned as witches,' noting it resurfaces every October. Citing modern historians, they put European witchcraft executions at roughly 40,000–60,000 over three centuries, 75–85% female; most victims were strangled or hanged. The author argues the inflated number undermines the real victims.

@mhdempsey frames Robinhood as a "rorschach test" for views on speculation, financialization, and whether the app has made equity investing more zero‑sum and power‑lawed, warning of margin compression and harmful marketing. He contrasts that critique with Robinhood CEO Vlad Tenev’s 2026-06-22 note announcing a proactive 10% workforce reduction with severance while claiming the business is strong.

Popular modernism: @bswud (citing @vpostrel) argues modernist architecture is more popular in California and Chicago than in Europe or the US East Coast because American postwar modernism often appears in premium private houses—especially in California—while UK postwar modernism was dominated by low-quality social housing. He points to The Priory, Cator Estate, Blackheath as a rare premium British example.

Meryl Streep faced typecasting at 40 when studios offered her three "witch" roles despite already owning two Oscars. In 1989 she chose the glamorous Mary Fisher in She-Devil, exposing her comic range and earning a Golden Globe comedy nomination; subsequent choices (Postcards from the Edge, Death Becomes Her) and a record 21 Oscar nominations (8 before 40, 13 after) reversed Hollywood's verdict over three decades.

mhdempsey recounts a relationship almost exactly a decade ago that was split between San Francisco and New York City; he made two prints of an image when his partner moved to NYC, kept his copy in work bags for the past ten years and only recently noticed it while unpacking. He later moved back to NYC, reflects that his life has been split, daydreams about connecting the cities, and ultimately says both are home.

Elon Musk and Whole Mars Catalog claim the accusation that he “killed millions” is baseless because no individual names have been produced. They argue Musk was only an advisor to the President, so institutional actors (President, Congress, courts) would have stopped any truly deadly spending cuts. They further assert critics targeted him to protect entrenched waste and fraud, despite his achievements in cheaper space launches and profitable electric cars.

Orca hunting culture is presented as the reason humans are rarely prey: the author cites a 1972 California case where an orca bit a surfer, then released him after a single taste (the surfer needed over 100 stitches). Orcas, the post argues, inherit narrow prey preferences across generations, use long-range echolocation and close-range visual matching to a learned search image, and abort when targets don't match—so no pod has ever learned to hunt people.