Personal AI News Digest
30 topics from 3 sources
4,093 words · ~20 min read
Anthropic moved its entire monorepo off Git to a closed-source system called Scion, and built a code review platform called Antique to replace GitHub's review features. Anthropic says coding is no longer the bottleneck. Merges went from about 2,500 code changes per day in May 2026 to over 30,000 per day by September, a roughly 12x increase in four months. The working copy of the monorepo is 25GB. CI alone performs about 25 million git checkouts per day. Around 3,000 staff each run 10-15 parallel Claude agents, which generates a large number of PRs and CI triggers.
Git's assumptions break down here. It was designed for clients that go offline and catch up later, but Anthropic's clients are nearly always online, so offline-oriented design adds inefficient checkouts and extra data transfer. Anthropic ran hundreds of git mirrors, but once the repo outgrew mirror disk capacity, it could no longer scale horizontally.
Scion's design is simple. Stateless-looking server nodes hold the whole repository, changesets are stored in Google Spanner, and large blobs are stored in Google Cloud Storage. Server and client are written in Rust. Antique's backend is Go and its frontend is React. Spanner runs with more replicas than the default three and uses multi-region replication, which trades write latency for durability. Full database backups, soft deletes on blob storage, and Spanner point-in-time recovery cover bad writes. Every component can autoscale, but it has not needed to, because the data is content-addressed and immutable, so caching does most of the work. A team of seven engineers (four on the core Scion server and client, three on code search, merge queues and related tooling) builds and runs it.
The migration used Claude heavily. 100% of the monorepo is now off Git, and a few smaller repos are migrating. Community resources such as skills, plugins and Claude Code's public repo remain on GitHub. Anthropic also launched OSS Scanner, which offers free security audits to open-source projects.
Anthropic did not open-source Scion. Nick G. said the reasons are bandwidth, since a no-support code drop is a lot of work, and that the code changes too often. Anthropic would rather share its design principles and work with GitHub, GitLab and the OSS community on scaling git, mainly around change sets, derived data and the checkout protocol.
Context: Google built Piper around 2011-2012 after outgrowing Perforce, and Meta built Sapling on top of Mercurial, later open-sourcing it along with its Mononoke server, written in Rust. Google was preparing for 10x more code in 2025, and Anthropic's growth is that scenario arriving early.
Attribution dispute: the creator of Mononoke publicly accused Anthropic of intentionally copying Mononoke's Bonsai data model and derived-data design, with personnel overlap. Anthropic originally gave no credit and added a small attribution only after the complaint. The author's view is that central VCS will stay niche, since few companies will rebuild source control, while a new batch of git-compatible vendors (Entire, Cursor Origin, Code Storage) targets AI-scale workloads.
Sources The Pragmatic Engineer
Links The Pulse: Anthropic builds its own source control system, engineering blog, deepdive published in July 2025, OSS Scanner, Perforce, Inside Google’s Engineering Culture, Sapling, documentation, open sourced Sapling, Monoke, wrote a post, Bonsai data model, a small attribution, Entire, Cursor Origin, Code Storage
Researchers disclosed two PHP-engine memory-corruption vulnerabilities, CVE-2026-109618 and CVE-2026-109619, named PHP2Shell. They sit in PHP's internal handling of out-of-memory conditions and fatal errors, not in a plugin or application auth logic. Researchers report pre-authentication RCE demonstrations against WordPress, Drupal, and Joomla, plus remote memory disclosure claims against any PHP application. Discovery is credited to multiple parties: OpenAI says Calif researchers used GPT-5.6 Sol under the Patch the Planet program, and a Wiz-affiliated account credits Gemini 4 Argon with discovery, exploit chains, detectors, and patch validation. OpenAI describes about two months of coordinated remediation with the PHP Foundation, maintainers, security providers, and governments, and a major PHP-engine change. Cloudflare deployed WAF mitigations for known vectors, free tier included, and called them a stopgap, directing operators to patch at 14:00 UTC. The NSA advised preparing for an immediate update and taking systems offline if timely mitigation was impossible. Reports say details leaked before the patch, which the supplied material does not confirm. Unverified claims include about 1.5 million servers running unpatched versions per Censys and several outages, none corroborated in the supplied material. Missing primary-source details include the exact fixed-version matrix, CVSS vectors, and the patch commit. OpenAI published a Codex skill to help operators check exposure. Commentators stress that WAFs create false confidence and that patching remains mandatory.
Sources AINews
Links [AINews] OpenAI’s Patch the Planet program is 1 of 2 parties to discover PHP2Sh…, Daybreak for Frontline Defenders, Patch the Planet, @sagitz_, @sagitz_, @TheRundownAI, @OpenAI, @itszn13, @dok2001, @NSACyber, @Eric_Wallace_, @gdb, @S1r1u5_, @fiveoutofnine, @SemiAnalysis_, @OpenAI
Anthropic opened a waitlist for Claude Duet, an Opus-5.5-powered proactive assistant. Reported initial access is for U.S. Pro and Max users. Examples include monitoring Slack, preparing questions, and making purchases after a request. Stripe describes Link-wallet purchasing and MCP-based invoicing, refunds, and revenue analysis. Separately, Anthropic's API version 2026-10-21 removes temperature, top_p, and top_k from Messages. New major SDKs default to it, while 2023-06-01 remains supported, so teams upgrading SDKs should review migration behavior. Reddit commenters noted Claude Duet is reportedly included in the existing $20/month plan, suggesting the proactive-agent layer is part of mainstream paid Claude usage. Critics argued that many assistant launches are thin chatbot wrappers over frontier models. One commenter argued the missing piece is structured personal-data integration and orchestration rather than model capability.
Sources AINews
Links Anthropic executive, access, payments integration, SDK defaults, version changes, Google Cloud distribution, Claude Duet: Anthropic just casually dropped the personal assistant calling it…, X/Twitter announcement from Claude
Recursive released Flint, an Apache-2.0 coding MoE with about 40B total parameters, 2.5B active, and a 262K-token context window. LM Studio says it needs about 24GB locally. Release claims are 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on SWE-bench Verified; these are not independently replicated. Advertised short-context speeds exceed 125 tokens/s on a MacBook Pro and 150 tokens/s on RTX Spark. A user measured 114–132 output tokens/s on an M5 Max with int4 MLX. Recursive says its automated Eureka Machine contributed sparse architecture choices, low-bit training, and kernels, claiming a 71% size reduction with almost no quality loss and development from scratch in under four months. vLLM reports 500 tokens/s on one GB200 at 100K context using NVFP4 without speculative decoding. OpenRouter offers free launch access.
Sources AINews
Links launch, specifications, results, local measurement, method, vLLM, OpenRouter, RecursiveSI/flint (40B A2.5B) · Hugging Face, https://blogs.nvidia.com/blog/local-ai-gtc-rtx-dgx-spark-station-sync/
Eagle-4 is a source-available speculative decoding stack using hardware- and workload-calibrated draft trees, with integrations for SGLang, vLLM, and TokenSpeed. Current benchmarks center on Qwen3.8-27B. Reported speedups over vanilla decoding are 2.22× for MTP, 2.57× for DSpark, 2.80× for DFlash 2, and 3.59× for EAGLE-4. The repo reports SPEED-Bench gains on 4× NVIDIA B300 of 3.42× on SGLang, 2.83× on vLLM, and 3.09× on TokenSpeed at single-request concurrency, shrinking at higher concurrency. Commenters asked for workload breakdowns, since averages can hide gains on structured outputs, and noted the draft model is about 3 GB larger than DFlash2. LiLiCorr, used by Nemotron Bolt, was suggested as a comparison.
Sources AINews
Links Eagle-4 Released: Faster than DFlash2 & DSpark, EAGLE-4, paper, https://github.com/syv-ai/HyperQwen
DeepSeek's released framework combines speculative decoding, dynamic prefill sizing, decode balancing by request count and KV length, and overlapping CPU scheduling with GPU execution. A reusable idea is adjusting speculative verification length from predicted draft acceptance and measured hardware throughput, so batches stay regular. A deployment-report summary cites 1.5M input tokens/s per prefill accelerator including cache hits, versus 60K excluding hits at a 96% cache-hit rate, with decode at 14K output tokens/s per accelerator. The same summary gives about $38 per accelerator-hour as revenue, not profit, excluding server, power, cooling, and staffing. SemiAnalysis says DeepInfer is absent from InferenceX because essential serving components are missing. Its profitability comparison was criticized for mixing hardware generations, peak and off-peak prices, and revenue versus profit. A commenter noted that the code may be tuned for DeepSeek's architecture.
Sources AINews
Links technical analysis, report summary, benchmark limitation, comparison critique, New from DeepSeek - DeepInfer -, image, GitHub repo, X post
Google open-sourced TorchTPU for training and serving through standard PyTorch tooling. SGLang's TPU plugin is reportedly powering production inference at Meta and supports several current open models. Its custom MegaKernel reaches a claimed 770 decode tokens/s for Muse Spark Open at batch size one on an 8K-input/1K-output workload. vLLM-TorchTPU adds prefix caching, speculative decoding, and disaggregated serving.
Sources AINews
NVIDIA's DSX POD packages AI infrastructure into containerized building blocks, with coverage claiming capacity expansion in days rather than months. These are product claims. Lambda demonstrated 19 HGX B200 nodes operating within the power budget of 16 using DSX MaxLPS. This is a proof of concept, not fleet-wide capacity already delivered.
Sources AINews
Links DSX POD, Lambda demonstration
OO-Harness uses specialized subclasses, execution-derived design lessons, and routing to raise AutomationBench from 61.6% to 77.4% with the same Gemini 3.5 Flash executor. Workspace-Bench rises from 25.5% to 39.5%. Agent S4, a meta-agent that builds tools from trajectories without success labels, reports 21.3% on OSWorld V2 with an open 27B model, 3.8× the same model with hand-built interfaces. ScienceBoard rises from 41.4% to 58.6%. These are paper-reported system results, not model-only comparisons. Separately, SuperMimic reports training across 129,514 motions on one GPU for three hours, versus a cited 128 GPUs for seven days, using a learned differentiable world-model surrogate with backpropagation through time. The >1,000× GPU-hour reduction applies to this motion-control setup, not robotics learning generally.
Sources AINews
Links paper summary, summary, results, mechanism
Mirendil says one researcher built its first base model in 3.5 months for under $10M in compute, including experiments and the final run. It discloses broad parameter ranges rather than exact configurations: 100B–500B for Base, 500B–2T for Large, and 2T–10T for XL. Independent evaluations, post-training results, and API access are still pending, so this is an unvalidated efficiency claim.
Sources AINews
Links announcement
Perplexity and NVIDIA propose JOLT, which addresses a self-distillation weakness where a privileged teacher exploits shortcuts unavailable to its student. JOLT trains the teacher with outcome rewards while penalizing excessive divergence from the student. The authors report matching GRPO on MATH with 6.5× fewer training tokens. JOLT+, which adds student outcome rewards, beats GRPO across five evaluated benchmarks.
Sources AINews
OpenAI reports that thousands of agents linked gene-expression results from millions of mouse-brain cells with literature in 32 minutes. One hypothesis connects DCTN1 disruption to impaired cholesterol transport, and another compares TSC1 and TSC2 disruption in cerebellar neurons. These are hypotheses for experimental follow-up, not validated treatments or drug targets.
Sources AINews
Links DCTN1 work, TSC2 follow-up
Artificial Analysis revised MLCR-AA's prompts, reference answers, and three-model judging panel. Sonnet 5.5 and GPT-6 Astra tie at 41.7%, and no model exceeds 42% under the combined completeness-and-accuracy requirement. Claude leads completeness, while OpenAI leads record-grounded accuracy. Revised scores are not directly comparable with the previous methodology.
Sources AINews
Links results, methodology, comparison
An independent investigation argues that SWE-bench Multimodal v2 does not meaningfully require images. A separate researcher argues that running the benchmark without images is the essential sanity check. This critiques what the benchmark measures; it does not prove that its underlying coding tasks lack value.
Sources AINews
Links investigation, ablation discussion
Cohere launched Rerank 5 in Pro and Fast tiers. Its opt-in Context Focus returns scored passages rather than whole documents, claiming up to 41% lower token consumption while maintaining accuracy. ParseBench reports NVIDIA's sub-1B Nemotron 3.5 Parse taking first place across its document-understanding benchmark.
Sources AINews
Links Cohere feature, availability, ParseBench
Ornith AI released Ornith-1.5-122B, a 122B MoE reasoning model fine-tuned from Ornith-1.0 with an RL self-improvement loop that jointly optimizes task generation, scaffolds, and rollouts. Artifacts include GGUF, unsloth GGUF, and NVFP4 variants. The model card claims 262K context with optional YaRN scaling toward about 1M. Commenters criticized the Qwen 3.5 base as older than newer alternatives, and estimated roughly 2× slower inference than A6B-active models because it uses A10B active parameters without sparse attention.
Sources AINews
Kyutai compressed its CPU-real-time voice-cloning TTS from about 100M parameters to 18.2M, 6.9M, and 1.9M variants. The method uses distillation from a roughly 300M teacher, baked-in classifier-free guidance, a PCA-derived 64-dimensional latent interface, structured pruning, and optional reward fine-tuning. Reported CPU throughput is about 9.9× real time for 18.2M and about 25× for 1.9M. The 18.2M model reportedly matches or slightly exceeds the original on WER and UTMOS. Users reported a profanity-filter artifact where obscenity was rendered as substitutes, which the team says it fixed. One commenter reported poor non-English output.
Sources AINews
Links Pocket TTS which supports voice cloning is now just 1.9 million parameters, Pocket TTS-small
fal first announced UI and agent access for Kling 4.0, with API support forthcoming. Kling later announced its official API rollout. Claimed improvements center on complex-motion stability, reference consistency, and cinematic control. The later announcement supersedes the earlier API-coming-soon status.
Sources AINews
Links fal launch, API rollout
Reuters reports that ByteDance, Alibaba, and Tencent placed over $35B in Nvidia chip orders. A Reddit thread cites a source saying Nvidia may fulfill only 5–10% of requested quantities and can produce at most 3 million RTX PRO 5500 cards globally in 2027, while Chinese customers requested over 6 million. The implied unit price is about $11,400. Commenters argued that the order book is supply-constrained and that high-margin datacenter demand may crowd out workstation and consumer cards.
Sources AINews
Links EXCLUSIVE: ByteDance, Alibaba, Tencent place over $35 billion worth of Nvidia c…, Reuters
GSMArena reports that Samsung has pre-sold about 70% of next year's general-purpose DRAM output and 100% of its HBM capacity under multi-year big-tech contracts. Expected effects include HBM price increases of 150–200% and DRAM price increases in double digits in Q1 2027. Commenters expect worse consumer and local-AI GPU affordability. One argued that the break-even point for local rigs versus cloud inference has moved, especially for MoE models that datacenters can multiplex across users.
Sources AINews
Links Be prepared, GPU price increase incoming --> Samsung sells out 100% of all 2027…, GSMArena reports
Reports say China is doubling RAM production, likely via domestic suppliers such as CXMT. Commenters argue that new fab capacity ramps over years, so the near-term effect is slower price inflation rather than cheaper RAM. CXMT targets 600,000 wafers/month by 2028, and other expansions include Samsung's new facilities and SK Hynix's Y2 and F17 investments. Output may mainly replace Chinese imports rather than flood export markets.
Sources AINews
Links When will we see these changes? --> China Is Doubling RAM Production to Keep Ev…
A kernel documentation change merged by Linus Torvalds removes the prior requirement to declare AI assistance with an Assisted-by tag on new code submissions. Commenters expect community pushback, while some contributors say the change may lower friction for upstreaming small hardware-specific fixes that otherwise remain local patches.
Sources AINews
Links As of 3 hours ago, the Linux Kernel no longer requires the usage of AI to be de…
A developer says their pipeline ingests about 100K articles per day from more than 40K sources in 40 languages, clusters them into events, and passed 11,079,667 articles. Claude Code reportedly queried aggregated cluster metadata rather than reading full articles, and the pipeline uses embeddings, heuristics, and an LLM judging step. Commenters questioned source-selection bias, aggregator ranking effects, and language labeling, and noted that full LLM analysis at this scale would be impractical.
Sources AINews
Links I handed Claude Code a database of 11 million news articles, it found some weir…, “The world right now” map
A user reports using Claude to diagnose a Samsung TV as Tizen rather than Android, then work around the lack of webcam access by serving a local page and capturing TV screenshots through a phone camera. The agent reportedly changed privacy and telemetry settings, deleted the advertising ID, disabled unused apps, and built a custom launcher and companion settings app. A separate commenter built a guest-oriented launcher integrated with Home Assistant and Plex. Another suggested the cleanyourtv GitHub project as a lower-effort alternative. A security-minded commenter warned that unvetted TV code could create a broad attack surface.
Sources AINews
Links I used Claude to fix my Samsung TV, QR screenshot, launcher, settings UI, I replaced my Samsung TV’s home screen with Claude Code, image, GitHub, write-up
Posts citing BBC reporting allege that Google employment contracts restrict public criticism indefinitely, discouraging discussion of AI safety and military uses. The supplied material includes no company response or legal determination. Separately, Anthropic published its annual Public Benefit Report and appointed Christopher Stubbs, an experimental physicist and former Harvard science dean, to its Long-Term Benefit Trust.
Sources AINews
Links report excerpts, researcher account, report, trust appointment
Waymo reports 271.3M rider-only miles with 0.01 serious-injury or fatal crashes per million miles, versus a modeled local human benchmark of 0.21, about 95% fewer serious crashes, and about 82% fewer injury crashes. An IIHS analysis of 2021–2024 federal crash data over about 50M driverless miles found 68% fewer police-reportable crashes and 81% fewer injury crashes, with an Austin result slightly worse than the human benchmark. Semafor reportedly says Waymo plans to scale from about 4,000 vehicles to 20,000 by end-2027. Critics note most exposure is dense urban, low-speed driving, and that the Austin sample of two crashes is too small for strong conclusions.
Sources AINews
Links Waymo’s robotaxis had 95% fewer serious crashes than humans. Now it plans a fle…, Waymo’s latest safety data
Unitree announced the H2-X Kung Fu Legend Guardian, a humanoid platform it frames as a serve-and-guard robot, claiming maximum endurance of about 8 hours. The post sits alongside other Chinese humanoid combat demos, such as Engine AI's T800. The source video was not accessible. Commenters focused mainly on safety and dual-use concerns.
Sources AINews
Links Unitree’s newest robot: H2-X Kung Fu Legend Guardian, X post, this writeup
A Reddit user reports that Fibo 2 Edit, a Hugging Face model, is fast and uncensored, with no observed color drift in simple edits, and attributes some quality to the Flux 2 VAE. ComfyUI support is in progress through a PR. Later reports say it struggles with complex prompts and multi-reference edits, and is faster than Qwen but not better. One report says source images are downscaled to about 1MP, causing blur.
Sources AINews
In this episode, Gergely Orosz talks with Will Larson (CTO of Imprint, former CTO at Carta and Calm, engineering leader at Stripe and Uber, author of The Engineering Executive's Primer, Staff Engineer, and Crafting Engineering Strategy) about how engineering management norms have changed since the 2010s, how AI is changing engineering roles, and what that means for the engineer-to-CTO path.
Digg's v4 rewrite: Digg was one of the top-10 most visited websites globally in the late 2000s and bigger than Reddit at the time. Larson worked there during the 2010 v4 rewrite, which was run to a hard deadline set by the returning CEO Kevin Rose. Digg was already losing traffic to Facebook, and a Google algorithm change caused a 60-70% traffic loss. The company had about 120 people and was burning cash. The v4 release alienated passionate users and had major performance and reliability problems. Larson and a small team fixed them; the root cause was Python memory leaks introduced by developers who did not know Python well. Larson's view is that Digg would have been overtaken by Facebook and Reddit regardless of the rewrite.
Yahoo incentives: Cofounder David Filo required approval for every server purchase to drive accountability. This produced a black market: one engineer had 30 servers left over from a cancelled project and traded them for things he needed. Servers were never returned, so Yahoo ran many outdated, power-hungry machines. Management practice was also thin then: in 2008-2010 Larson had only two 1:1s with his manager, one about someone else's performance and one when he handed in his notice. Annual or rare 1:1s were common.
Staff+ and business focus: At Stripe, Larson led about 250 people, and many senior engineers sought more 'engaging' problems. He was seen as violating an unwritten rule by not supplying interesting enough problems. Today, putting business problems first is a baseline expectation at Staff+ levels.
ZIRP and hiring: Larson attributes many 2010s management practices to the zero interest rate period (ZIRP), and recommends Ranjan Roy's 'ZIRP explains the world'. In that era 'hire more engineers' was the default answer for engineering leaders; Larson says it no longer is. Companies also often kept the highest-paid engineers off on-call to help hiring and retention, but on-call ownership leads to more reliable software because teams feel the pain of their own outages. Gergely wonders whether AI agents could auto-mitigate expected incidents.
AI and flow: AI has made engineers more productive but less satisfied. Larson knows someone who retired from tech because AI destroyed their ability to stay 'in the flow'. Larson says focusing on outcomes is what keeps him enjoying the work.
Tech lead managers: The TLM role is in high demand. Engineering managers are now handling 12-14 people instead of 4-8, and directors manage more TLMs. Larson calls TLMs a career trap in his writing but says they suit builders, and some of his best Imprint hires are managers who want to keep writing software. A great TLM has the capability but no desire for bureaucratic work.
Technical decisions: The old advice that managers should let engineers make technical decisions is outdated. More companies now expect managers to be hands-on and in the details. Larson argues that great engineering executives must be both technologists and bureaucrats, and that many past executives from the middle-manager path were bureaucrats first.
Retire or re-skill: Larson moved from leading 250 people at Carta to leading 25 at Imprint on purpose, because he believes the engineering leadership profession as we know it may not exist in three years. He argues line management is more engaging now because tools for line management and for writing software are very good, while bureaucracy tools remain poor, which is why talented engineers get jaded. His advice to middle managers is to go back to line management and re-skill, or retire.
Career ladders still matter, and Larson walks through his path to CTO, his early career, and his time at Uber, Stripe, Carta, and Calm. Imprint is hiring several engineering manager roles to rethink how software teams work with agents.
Sources The Pragmatic Engineer
Links The revised rules of engineer…, Listen now, The revised rules of engineering management with Will Larson, chapters, The Pragmatic Engineer, Will Larson, Imprint, hiring for several key engineering manager roles at Imprint, ZIRP explains the world, https://en.wikipedia.org/wiki/Digg, on his blog, is a career trap, Engineering leadership skillset overlaps, The Engineering Executive’s Primer, Staff Engineer, An Elegant Puzzle, Crafting Engineering Strategy, Getting an Engineering Executive Job, What is a Staff+ Engineer?, Inside Stripe’s engineering culture, Staying technical as an engineering manager, The end of ZIRP and what it means for startups and the tech industry
Anekha Sokhal, CEO of Moshi, discusses her path from a maths, operational research, statistics and economics degree at Warwick to energy derivatives trading at JP Morgan, then to founding Moshi. She won a Fulbright data scholarship (one awarded in the UK that year), did ML research published at ICLR with Stability AI, and worked with NASA. Her framing is that Moshi is the program she wished she had as a trader, with the pitch 'if Bloomberg, Claude and Netflix had a baby, it would be Moshi.'
Moshi provides real-time insights for financial traders, starting with commodities, energy and macro. She says it processes about 300 new items in the time it takes to say one sentence (around four seconds). The inputs are structured time series (futures prices and technical indicators), fundamentals time series (supply and demand), unstructured data (news, X posts, podcast audio), and private user data such as analyst subscriptions and internal company feeds. Pipeline: data cleaning and transformation, statistical and ML tagging, an ontology layer, and an n-dimensional graph that outputs a software feed pushing portfolio-relevant updates to the user. Users can also query it and get answers in about 30 seconds that would otherwise take days.
The recommendation logic is built on the graph so traders see second and third-order effects, for example an energy equities company action that indirectly moves oil, which a news stream would not connect. She argues recommender systems are underrated in AI engineering and that a dashboard is not enough: adoption is a UI problem, so the thesis is that solving the recommendation and UI problems together is what makes tools stick. Swyx notes the host's own experience at Balyasny buying the X firehose for roughly $300,000 a month, and the noise problem with social data.
On source trust, users follow specific industry voices such as suppliers and LNG facilities. Moshi filters sources and shows a confidence score alongside what the wider conversation says. Divergence between sources (Reuters, Bloomberg, WSJ disagreeing) is treated as a signal rather than noise. For documents, Moshi uses an in-house vision-language model to extract data from messy PDFs instead of sending customer data to third-party services like Reducto. Build vs buy depends on the team: most of hers has NASA-adjacent experience, so building in-house is practical for them, while a smaller team may reasonably buy.
At AI Engineer New York, the audience was mostly traders at funds and consultancies, plus AI engineers under pressure from management who say 'you are the team' and expect one person to build a system usually built by about 10 people. Inbound led to back-to-back meetings and an extended stay in New York plus Houston meetings. Anekha says the priority is meeting clients over hiring. One job posting drew 2,500 applicants in four days, mostly noise including visa-constrained and undercover candidates. She proposes round-table sessions with several prospects at once, as she experienced on the buy side, and a webinar to handle inbound. The episode closes with swyx's path into AI and the AI Engineer conference.
Sources Latent.Space
Links Latent Space: The AI Engineer…, The Recommendation Engine for…, chapters, Listen now, The Recommendation Engine for Commodity Traders — Anekha Sokhal, Moshi, Latent.Space, Watch Anekha Sokhal on YouTube, Moshi, Anekha Sokhal, GIQ research paper, Reducto, AI Engineer New York, Latent.Space