We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Coverage is incomplete: some monitored sources or documents could not be processed. This brief covers the available verified material.
1. Funding & Deals
Agent-runtime infrastructure is becoming an acquisition category. Baseten announced that it acquired Blaxel, which builds sandboxes for agents running code and background tasks. The press release did not disclose a price, but Newcomer reports it heard roughly $300 million; Blaxel had raised only $7.3 million in seed funding from First Round Capital, Liquid 2 Ventures, and Y Combinator. The signal is a strong markup for infrastructure that inference providers could have chosen to build internally.
Power infrastructure is moving into the AI stack at seed stage. Blue Core Energy raised a $50 million seed round for small nuclear reactors mounted on barges; the episode explicitly connects the deal activity to data-center and AI power demand. The company had two barges in the water and was pursuing design certification through the Nuclear Regulatory Commission, Coast Guard, and Department of Transportation. Portability, safety, and regulatory approval remain the diligence gates.
Replit is showing a platform-to-company-to-acquisition loop. The two-person Test 13 team says it generated more than $700,000 in 18 months, grew more than 400% year over year, and offered local customers software services at 60–80% lower prices; Replit then acquired the business. Replit CEO Amjad Masad framed it as potentially the first of many companies built on the platform.
2. Emerging Teams
A self-reported autonomous-AI-scientist project is becoming a compute-heavy team. Suhail’s public thread describes an autonomous AI scientist, a completed seed round, growth from one person to three, acquisition of 64 B300s, additional compute being locked down, and a latest update that a “critical third hire” was made. These are meaningful formation and infrastructure milestones, but the current evidence is operational rather than a validated research result; diligence should focus on research throughput per unit of compute.
Mass Magnetics (YC S26) is targeting a materials bottleneck for robotics and defense. YC says the company recycles magnets from EV motors into new rare-earth magnets, avoiding much of the capital-intensive refining required by mining; the pitch notes that 90% of the material is conventionally discarded. The early-stage thesis is circular supply for a strategically constrained input, rather than another software layer on robotics.
Small control-plane products are forming around agent failure modes. Guardrail by NEAT is a local proxy for coding agents and LLM APIs with hard session/project caps, local prompt and completion handling, and attribution to loops, agents, or keys. Its own caveat is important: it blocks the next request rather than interrupting an in-flight generation, and sub-agent-level attribution is still a gap. Pulse takes a complementary approach for MCP servers, offering usage, failure, and latency telemetry without collecting prompts or tool inputs/outputs; tracking is asynchronous, but its exact overhead is still being benchmarked. Together they point to an investable control layer built around spend, observability, and privacy—not model novelty.
3. AI & Tech Breakthroughs
The Navier–Stokes episode is increasingly a cost-curve story.Clouded Judgement reports that OpenAI used 10,000 agents for three to four days to produce a more-than-100-page manuscript and formalized proof; the effort reportedly generated about 2.7 million messages and 130 billion output tokens, with outside cost estimates of $10–40 million versus OpenAI’s description of “millions of dollars” in compute. The article’s more useful investment point is the slope: it cites estimated 9×–900× annual declines in the cost of reaching a given benchmark and argues that today’s low AI gross margins are point-in-time measurements on a falling curve. The source also notes allegations that the work drew on unpublished research, so provenance remains a material caveat.
AlphaGenome Atlas turns genome-scale prediction into a prioritization tool. The Atlas is described as a one-petabyte database of predicted molecular effects across the human genome, including its 98% non-protein-coding regions and roughly nine billion possible single-letter changes. It helped identify a variant predicted to disrupt RNA splicing in an unresolved rare-disease case, but is explicitly positioned as a way to decide where experiments should focus—not as a replacement for experiments.
Document AI is adding uncertainty and provenance to extraction. LlamaParse’s high-effort mode provides page-level confidence scores, explanations, and an additional check against the original document, priced at five extra credits per page; LlamaIndex says the scores can trigger human review or automated fallback logic. Reducto’s r-1 combines layout detection, reading order, tables, formatting, grounding, and granular citations in one request, while claiming up to 20% fewer parsing errors and a one-cent-per-page all-in cost. The shift is from “did the parser return text?” to “which outputs are safe for an agent to act on?”
4. Market Signals
Sovereign AI is being financed as a full stack, not just a local model. Mistral’s $3 billion Series D is a late-stage signal rather than a seed/Series A comp, but the strategy is directly relevant to early-stage infrastructure: investors describe the company as evolving from a frontier lab into an enterprise platform with open models, forward-deployed engineers, and its own compute. The stated use of capital includes compute, infrastructure, commercial expansion, international growth, and a full-stack AI cloud. Open-weight models are valuable to regulated buyers because they are inspectable, portable, and deployable on customer infrastructure; the interview frames Mistral as an option between increasingly capable Chinese open models and American closed models.
The business-model risk is equally important. Newcomer describes Mistral’s recent shift toward institutional deployments and hosting third-party open-source models, and quotes a VC warning that it could become a European services company rather than a frontier-model competitor. At the cheaper end of the stack, Smaug Flash is being marketed as an open-weights fine-tune for personal agents with listed prices of $0.10M input and $0.40M output; its claim of “DeepSeek Flash-level performance at 300% cheaper” is promotional, but the weights and model card are public.
The largest early-stage whitespace is workflow integration, not AI awareness. An Andrew Ng/Jad Masad discussion cites a Goldman Sachs survey in which 76% of small businesses already use AI but only 14% have built it into core operations; OECD data cited later puts core-activity use at 29% even among small businesses already using generative AI. The video’s operating test is narrow and demanding: choose one workflow, run a pilot for 90 days, and verify payment, repeatability, and faster delivery for the second customer. This aligns with a16z’s estimate that the median U.S. company spends $12 per employee per month on AI versus $7,000 for the top 1% of its dataset, with limited diffusion beyond coding.
Independent evaluation is becoming a governance and market requirement. Joe Benton says he left Anthropic’s safety team because AI companies are underinvesting in safety and joined METR to conduct independent evaluations. He calls for disclosure of recursive-self-improvement progress, safety incidents and near-misses, minimum standards, and independent guarantees. For investors, the implication is that evals, incident reporting, and enforceable controls are part of frontier-AI diligence rather than an optional research layer.
Cybersecurity is emerging as a model-company revenue lane, with a backlog caveat.Newcomer reports that top-tier models are effective at finding, patching, or exploiting vulnerabilities; Modal uses them as a partial replacement for external security consultants, while Town says continuous monitoring has become economically worthwhile. The counterpoint is that much of the demand may come from cleaning up legacy systems and could decline after those vulnerabilities are fixed, although the article also reports a broader shift in budgets toward AI security tools.
5. Worth Your Time
- Watch — Why Mistral Just Raised $3B to Build Sovereign AI. The strongest segment explains the move from open-weight research lab to enterprise platform, forward-deployed adoption, and customer-controlled compute.
- Watch — The Biggest AI Opportunities Are Hiding in Boring Businesses. A practical framework for finding workflow-specific wedges where domain knowledge, customer access, and a narrow paid pilot matter more than coding speed.
- Watch — AI researchers go full doomer while Apple puts AI front-and-center. Useful for separating genuine control concerns from incentive-driven “capability flexing,” while keeping nearer-term labor, environmental, and governance risks in view.
- Read — Clouded Judgement: Paying for the Curve. The clearest short treatment in the period of why first-wave AI costs and margins may be less informative than the rate at which capability costs are falling.
- OpenAI reported a multi-agent approach to the Navier–Stokes problem: the system used 10,000 agents and 130 billion output tokens . The discussion framed this as compute- and orchestration-driven leverage—large-scale brute-force work rather than a singular, inexplicable insight .
- Enterprise AI sovereignty is emerging as an investable infrastructure theme. OpenAI said it could not rule out that deidentified usage data from mathematicians helped improve its models . The episode warned that zero-data-retention is only best-efforts and not guaranteed , while proposing self-controlled hardware, models, or private VPC deployments for sensitive workloads; it also said AI application companies were moving off frontier models because of these concerns .
- Frontier-model platform risk threatens application-layer defensibility: the panel said model providers reserve the right to enter vertical applications, citing Anthropic’s Claude Code and Claude Design launches and Cursor’s reported feeling that it had been “rugpulled” . This is a caution for investors underwriting startups built primarily on another lab’s model API .
- Autonomy claims should be stage-gated. The discussion distinguished “prosaic RSI”—AI assisting researchers, including writing much of their code—from “RSI maximalism,” in which AI independently designs and launches successive training runs . The former was described as observable, while the latter remains a much larger leap with many intermediate intervention points .
- Open-weight regulation is a concentration risk, according to the panelists: they argued open-source AI can reduce costs by 50x , while central monitoring and rollback requirements could effectively prevent open weights from being published or hosted and reinforce a closed-model duopoly .
- Agentic-model architecture and security: The episode describes a frontier model whose inference loops feed outputs back into another model, enabling longer autonomous computer-use runs while making chain-of-thought and interpretability harder to monitor. The discussion shifts the control point toward environment-level permissions, access controls, governance, zero-trust, and just-in-time access, identifying potential nonlinear demand for cybersecurity startups in this layer. Popular open-source tools such as Blender could also become default skill interfaces for models, concentrating value around widely adopted projects.
- Agentic assistants and startup-versus-incumbent dynamics: The discussion reports that Instinct can absorb email and administrative work and handle tasks such as booking flights and hotels, while Meta is pursuing a competing assistant; startup speed and risk appetite are presented as first-mover advantages against incumbents’ distribution, compute, and infrastructure advantages. At scale, these assistants require efficient, fast inference and high uptime, may specialize by task, and face unresolved permissioning and privacy-segmentation issues between work and personal contexts.
- World models and serving infrastructure: The episode reports that ByteDance is developing a real-time world model and treats open distribution, China’s ecosystem push, and ByteDance’s consumer-data access as signals of strategic importance; it argues that real-time video is approaching useful latency and fidelity for content, marketing, and robotic simulation. This creates a serving-layer opportunity: the models are inference-intensive and costly, while Reactor is described as making open world and video models efficient and performant through a developer API; current memory and KV-cache limits can still break scene consistency. The emerging market structure is described as a mix of closed frontier APIs and open-weight models that companies can serve to own their intelligence end to end.
Blue Core Energy: Raised a $50 million seed round to develop small nuclear reactors mounted on barges, a model linked to investor interest in power infrastructure for data centers and AI. The startup emerged from stealth in July with two barges already in the water and is pursuing design certification through the Nuclear Regulatory Commission, Coast Guard, and Department of Transportation; portability and safety remain open questions.
Stoke Space: Raised $1 billion, bringing total funding to $2.3 billion, in a round led by 72 Ventures, Steve Cohen’s fund. Its technical bet is a fully reusable launch vehicle: unlike SpaceX’s current Falcon 9 model, Stoke aims to reuse both rocket stages. The episode sees a market opening as SpaceX shifts toward Starship and away from Falcon 9/Falcon Heavy, while Starship’s upper stage is described as oriented mainly to Starlink, potentially leaving demand for alternative launch providers. Stoke’s CEO says the first launch is targeted for early 2027.
AI safety and governance: Anthropic researcher Jacob Coxin resigned after warning that the company’s AI development could end humanity; Anthropic’s alignment lead amplified the warning, writing, “We really do earnestly believe AI could kill all humans.” The panel argues that major AI labs may not fully control their models, that doomer/AGI framing can obscure nearer-term labor and environmental harms, and that investment- and growth-driven company structures make slowing development difficult.
- AI-driven venture outcomes are reshaping allocator math. a16z’s accompanying post claims the “SpaceXAI IPO” produced more exit value than the previous five years of venture exits combined, reporting $2.18T in 2026 YTD exits versus $75B raised in 2025 and $222B in 2022. The linked thesis is that machine-intelligence capability, autonomy, and cost are improving exponentially, with robotics still ahead.
- Private-market momentum is bifurcating. Median time between rounds for actively raising unicorns fell to one year in Q1 2026 from 1.5 years in 2024, while 51.2% of private companies that had cleared unicorn status had not raised in more than two years.
- AI infrastructure is an emerging financing theme, while legacy software is a caution. a16z says private credit is moving up-stack to underwrite the AI-infrastructure buildout, but warns that earlier loans to PE-backed software may face write-offs as AI compresses multiples, growth slows, rates rise, and refinancing approaches.
- Liquidity and fund selection remain central risks. The article flags LP concerns over limited DPI and opaque venture marks. Its cited PitchBook screen of 2,143 global VC funds found that only 17.0% returned at least 2x invested capital, 6.7% reached 3x, and 2.4% reached 5x, supporting a highly access- and manager-concentrated venture market.
- Mistral funding and scale-up: Mistral raised a $3 billion Series D, described as the largest equity round ever for a European technology company. The capital is aimed at scaling compute and infrastructure, accelerating commercial and product growth, expanding internationally, and building a full-stack AI cloud; Samsung and EQT joined the cap table and the ASML partnership deepened.
- Founding-team pedigree: Investors described Mistral’s founders as part of a small cohort with experience training foundation models at scale, highlighting Llama-side experience, Timote’s results, and Arthur’s Google expertise. The team has rapidly evolved from research into engineering, product development, and commercialization.
- Open-weight and sovereign-AI thesis: Mistral’s models are positioned as inspectable, portable, deployable on customer infrastructure, and customizable, small, efficient, and multilingual. The strategy targets regulated and mission-critical buyers seeking model control, provenance, privacy, and cost efficiency, with Mistral positioned between increasingly capable Chinese open models and American closed models.
- Compute-infrastructure investment signal: Mistral’s compute strategy spans training its own frontier models, supplying capacity to customers for adoption, customization, and privacy, and providing infrastructure to hyperscalers; Microsoft is expected to use its expanded European GPU infrastructure. The thesis is that European customers in finance, defense, manufacturing, and the public sector need alternatives to U.S. cloud dependence.
- Lightspeed’s representatives said the firm participated in every Mistral financing round and remains one of the company’s largest shareholders, signaling sustained investor conviction.
- Scott Kupor endorsed the thesis that ignoring venture’s power law is dangerous.
- The linked analysis argues that machine intelligence is already producing exponential improvements in capability, autonomy, and cost, before robotics has meaningfully arrived, making high-ambiguity, high-upside technology a core venture opportunity.
- Private-market activity is bifurcating: the median time between rounds for actively raising unicorns fell to one year in Q1 2026 from 1.5 years in 2024, while 51.2% of companies that had previously cleared unicorn status had not raised in more than two years.
- Manager selection and access are presented as more important than broad diversification: the article says outcomes are concentrated among a small group of funds whose winning investments generate follow-on rights, founder referrals, and information advantages; it cites only 20 of 3,000 U.S. VC firms as achieving consistent 3x net returns over two decades, while a PitchBook screen found 6.7% of funds reached 3x and 2.4% reached 5x.
- Liquidity is the main counterweight: LPs may be “knee deep in TVPI” but short on DPI; secondaries can provide liquidity but may force sales of power-law winners, while AI-driven multiple compression is flagged as a risk for legacy software and private-credit portfolios.
- Vertical AI agents are an emerging opportunity in overlooked SMB workflows. The discussion highlights end-to-end property-management agents and ranks property management, freight/logistics, legal-document workflows, and dental/medical billing among the more valuable but harder verticals, where delays, document handling, rejected claims, and manual follow-up directly destroy revenue.
- AI adoption leaves substantial core-operations whitespace. The video cites a Goldman Sachs survey in which 76% of small businesses used AI but only 14% had integrated it into core operations; it also cites OECD data showing traditional and physically intensive sectors lag information and communications, with only 29% of small businesses already using generative AI applying it to core activities.
- The moat is shifting from coding speed to domain expertise, customer access, and product judgment. Jad Masad emphasizes embedding tacit domain knowledge into agents, while Andrew Ng describes deciding what to build as the “product management bottleneck” after building costs fell. The source therefore favors founding teams with lived workflow expertise and customer relationships, validated through a narrow paid pilot whose result is repeatable within 90 days; high-stakes areas still require domain accountability, with legal judgment remaining with lawyers.
- Demis Hassabis brings unusually deep technical and founder pedigree: he was a chess master at 13, wrote a commercially successful computer game at 17, studied at Cambridge, completed a doctorate on memory, and then started DeepMind.
- His founding experience blends AI research with product and team execution: at Elixir Studios, Hassabis used game development to fund early AI research with researchers including Dave Silver, while learning to manage startups and creative engineering teams; he later carried that multidisciplinary model into DeepMind.
- The strongest investment thesis is AI for science and medicine: the talk credits AlphaFold with determining the shapes of 262 million proteins and giving the results freely to researchers, while Hassabis identifies Isomorphic Labs as a DeepMind spinout focused on accelerating drug discovery.
- A second emerging theme is AI-augmented creative tooling rather than simple replacement: DeepMind has worked with leading artists, musicians, game designers, and filmmakers to shape tools that let professionals prototype more ideas, reduce production costs, and help emerging creators demonstrate projects earlier.
- Hassabis frames full AGI as potentially only a few years away and warns that technical safety, economic distribution, and societal purpose must be addressed as systems become more autonomous.
David Ulevitch cautioned that Flock’s visible engagement should not be treated as equivalent to U.S. audience size, writing: “People see 50k likes and think it’s 50k Americans. It’s not.” The referenced example estimated that 76% of engagement on Coxon’s post came from outside the U.S., based on a sample of 3,500 reposts, mainly from India, Indonesia, and Mexico.
Martin Casado highlighted a report alleging that Chinese “transit stations” or AI gateways bypass account and VPN restrictions by pooling inexpensive or geo-arbitraged subscriptions, exposing them through APIs, and routing users across usage limits. The report further alleges that these gateways may store and sell usage traces as training data; it says a researcher purchased such data and extracted user credentials. It also claims some gateways advertise access to premium models such as Claude Opus while routing requests to cheaper models including Kimi, DeepSeek, and GLM.
- OpenAI announced that roughly 10,000 AI agents worked for about 88 hours on a proposed Navier–Stokes solution, followed by 17 hours of formalization and verification in Lean; the construction addresses the forced version of the problem. The report also notes that Anthropic accused OpenAI of using Codex conversation logs, a charge Sam Altman denied.
- Google DeepMind released AlphaGenome Atlas, a one-petabyte database predicting molecular effects across the human genome, including its 98% non-protein-coding regions and roughly nine billion possible single-letter DNA changes. The Atlas helped identify a variant predicted to disrupt RNA splicing in an unresolved rare-disease case, but is positioned as a prioritization tool rather than a replacement for experiments.
- Hermeus unveiled Ramjet-X, an expendable high-speed test vehicle released from its reusable Quarterhorse aircraft; Quarterhorse provides the initial speed and altitude, then returns for another mission while Ramjet-X continues under its own power. Hermeus plans to sell this as “Flight Test as a Service” for sensors, materials, autonomy, communications, and propulsion testing, with integrated vehicle testing planned for 2027.
- Reducto says its document-ingestion platform processes more than one billion pages per month for AI teams including Scale AI, Harvey, Toast, and Vanta. Its r-1 parsing model combines layout detection, reading order, tables, formatting, grounding, and granular citations in one request, with claimed reductions of up to 20% in parsing errors, improved high-volume latency, and an all-in cost of $0.01 per page.
- Neko Health, founded by Hjalmar Nilsonne and Spotify founder/CEO Daniel Ek, launched a New York location after expanding in the UK and Stockholm. Its $499 preventative-health visit combines full-body scanning, bloodwork, cuffs, and laser/light measurements, with the stated goal of delivering increasingly comprehensive screening at that price.
Mass Magnetics (YC S26) recycles magnets from EV motors into new magnets for robotics and defense, avoiding much of the capital-intensive refining associated with mining. YC says rare-earth magnets are present in nearly every electric motor and that 90% of the material is conventionally discarded.
- AI adoption remains highly uneven: David George cites median U.S. company spending of $12 per employee per month versus $7,000 for the top 1% in one dataset; adoption beyond coding is still limited, while leading-edge banks spend roughly 1% of headcount cost on AI tools. This suggests substantial enterprise expansion runway, but also that current value capture is concentrated among the most advanced adopters.
- The discussion frames AI economics as a “compute vending machine,” where spending directly buys compute that can improve the product, and describes returns as unusually concentrated; it contrasts AI reaching $100B in four years with SaaS taking 15. The speakers characterize the market as winner-take-most: “the category winner takes it, second place takes scraps.”
- Venture returns are sharply concentrated: Accolade’s analysis of 3,000 U.S. venture firms found only 20 delivered consistent 3x net returns over two decades, with access to category-defining companies identified as the common factor; the discussion also flags mid-sized venture as being squeezed.
- Method and Palantir announced the Cardinal Program, a consortium of AI, infrastructure, and security partners focused on national cyber resilience. The program will provide select municipalities, utilities, and critical-infrastructure operators with free, opt-in, continuous autonomous security assessments supported by expert operators and human oversight, responding to cyber adversaries using AI.
- Joe Benton says he left Anthropic’s safety team and is joining METR_Evals to conduct independent evaluations; he argues AI companies are underinvesting in safety and calls for disclosure of recursive self-improvement progress, safety incidents and near-misses, minimum safety standards, and independent guarantees.
- Martin Casado responded, “DoE!! It’s time! Absolutely the most reliable way to ensure a pause,” signaling support for a pause mechanism amid the frontier-AI safety concerns described in the quoted post.
- Venture returns are exceptionally concentrated: a16z says the top 1% of VC investors capture 57% of industry net profits, while the top 5% capture 90%; among 2,143 global VC funds with 2000–2018 vintages and reported DPI, only 17.0% returned at least 2x invested capital, 6.7% reached 3x, and 2.4% reached 5x.
- Private-market momentum is bifurcating: the median time between rounds for actively raising unicorns compressed to one year in Q1 2026 from 1.5 years in 2024, while 51.2% of companies that had previously reached unicorn status had not raised for more than two years.
- The article frames machine intelligence as a major venture opportunity, citing exponential gains in capability, autonomy, and cost before robotics has broadly arrived; it argues venture’s power-law structure can still generate outperformance even if most portfolio companies fail.
- Liquidity and AI-driven repricing are key risks: venture investors still face limited DPI and opaque marks, while AI-driven multiple compression and slower growth may cause write-offs in earlier loans to private-equity-backed software; private credit is simultaneously moving toward financing the AI-infrastructure buildout.
- Martin Casado contrasts the Internet’s history of major malware incidents—including Morris, Michelangelo, Melissa, ILOVEYOU, and CodeRed—with what he describes as a remarkably small number of security-significant AI events despite extensive effort and spending, alongside a relative lack of security sophistication among AI developers.
- AI investment thesis: An a16z-highlighted discussion frames AI as turning spending into a “vending machine”: capital buys compute, which can directly improve the product. The panel argues AI is sold against labor budgets rather than software budgets, with category winners capturing most of the value and second-place companies receiving “scraps.”
- Funding-market bifurcation: Among actively raising unicorns, the median time between rounds fell to one year in Q1 2026 from 1.5 years in 2024, while 51.2% of private companies that had reached unicorn status had not raised for more than two years.
- Access and liquidity are major selection constraints: Accolade’s review of 3,000 U.S. venture firms found only 20 had delivered consistent 3x net returns over two decades, with access to category-defining companies the common factor. The article also identifies liquidity as venture’s central critique, noting that secondary capital remains small and concentrated in a narrow subset of companies.
- Open models are positioned as a complementary enterprise layer alongside stronger closed models, enabling customized agentic workflows while remaining in perpetual performance catch-up.
- The open–closed performance gap is described as roughly 4–6 months, with leading open models coming from Chinese labs since around 2024.
- Adoption is shifting toward Chinese open models for cost reasons: Perplexity rapidly adopted DeepSeek R1, while Thomson Reuters is building on Qwen to rely less on Anthropic. This usage has also triggered regulatory scrutiny of DoorDash, Airbnb, Anysphere/Cursor, and Apple.
- Distillation is identified as the central open-model debate in 2026; a cited paper reports reasoning-trace extraction from proprietary APIs, and Anthropic confirmed the technique was used by Chinese labs. Interconnects rejects the claim that distillation alone explains Chinese models’ proximity to the frontier.
- AI is strengthening the venture power-law thesis: a16z argues that machine intelligence is already producing exponential gains in capability, autonomy, and cost before robotics arrives, making venture’s asymmetric-return model especially relevant because a small number of winners can offset most portfolio losses.
- Private-market momentum is bifurcating: the median time between rounds for actively raising unicorns fell to one year in Q1 2026 from 1.5 years in 2024, while 51.2% of companies that had previously reached unicorn status had not raised for more than two years. a16z also argues that access is concentrated among a small group of funds: winning investments provide follow-on rights, founder referrals, and information advantages, while a PitchBook screen found only 17.0% of qualifying funds had returned at least 2x invested capital, 6.7% had reached 3x, and 2.4% had reached 5x.
- Cautionary signal for incumbent software and credit: a16z warns that AI-driven multiple compression, slower growth, higher rates, and refinancing needs may lead to write-offs among earlier loans to PE-backed software companies; unlike venture, these investments generally lack power-law upside to offset losses.