levelsio calls out AI firms' hype over 'software is dead'
levelsio notes that AI company employees jumped on the 'software is dead' debate, claiming it's the best time ever to build a startup, arguing the sentiment is somewhat dishonest.
2026-10-13
All AI Pulse updates published on 2026-10-13.
51 signalsHarbor offers coverage for AI agent decision risks
Agents can now approve, price and pay on their own, and companies bear the cost of bad decisions. Harbor proposes a Connect-Measure-Block-Insure flow covering every agent a company runs.
Frontier labs say unpaid entry still too costly
Financially independent people asked about joining frontier labs without salary. The labs' response: each person's GPU and token consumption far exceeds a salary, and wasted resources are unaffordable.
RouterPlus pools team demand to unlock inference discounts
Inference pricing favors the biggest volume commitments. RouterPlus groups many teams' demand into one large provider deal, so each team pays the bulk rate on its share of traffic.
Google Play opens 'personal apps' for AI Studio builds
On October 13, Google announced that Android apps handcrafted with AI Studio can be installed directly on your own phone, no developer registration required.
Grep ships one new feature daily, starting with Custom Skills
Following its YC launch, Grep is releasing one feature per day this week to automate repetitive work. Day 1 brings Custom Skills, letting users build tailored skills in Grep.
LangChain Academy is almost in session
LangChain announced that LangChain Academy, its learning program for developers, is almost in session.
Study: division of labor helps self-improving agents avoid errors
New research suggests dividing labor among self-improving AI agents can prevent them from reinforcing their own errors, an idea linked to open-ended search in metascience.
Hassabis praises Gemini 4 Argon for knowledge work and coding
Demis Hassabis says Gemini 4 Argon excels at knowledge work and coding. His team built a wizard chess game with it, showcasing its physics understanding and 3D modeling.
Tesla calls FSD Supervised a superhuman guardian angel
Tesla described FSD Supervised as a “superhuman guardian angel,” framing the driver-assistance system as an added safety layer.
Gavin Baker calls AI race highest-stakes corporate chess ever played
a16z quotes Gavin Baker calling AI a record-stakes corporate chess game, crediting Meta’s comeback and noting Gemini’s rise a year ago.
Travel agent predicted a delayed flight and offered backups
A user said their travel agent tracked the plane’s prior leg, predicted a delay, offered last-minute alternatives, and arranged rental-counter flexibility.
Agent traffic strains inference servers with repeated prefixes
Rohan Paul notes agent turns resend long near-identical prefixes and request short tool calls; Inferact’s Yifan Qiao discussed this Serve topic at DigitalOcean’s summit.
Suno cover flips “Oh, Viburnum Blooms” narrative roles
A creator shared Suno’s cover of “Oh, Viburnum Blooms”: the old song had the girl suffering, the new version the boy, released in June 2026.
New Skill adds subtitles, memes, music to screen recordings
The author shared a Skill that adds bilingual subtitles, highlighted text, memes, and music to recordings, makes covers and end credits, and mentions Claude and Codex.
Depot CTO says machines are becoming our users
AWS Startups quotes Depot CTO Jacob Gillespie saying machines are becoming users, adding AWS helped Depot build for both humans and AI in the developer space.
IMF says AI lifts growth but erases entry-level jobs
Business Insider cites IMF chief Kristalina Georgieva saying AI props up growth while erasing young workers’ first jobs; IMF research estimates AI may affect about 60% of jobs.
MagicOS bets the canvas stays with designers
alexdanilowicz launched MagicOS, arguing designers are not being replaced; while many tools give agents their own canvas, MagicOS keeps the canvas with the designer.
🌱 Can AI help farmers save water?
🌱 Can AI help farmers save water? I interviewed twin brothers Edgar and Eduardo Bello Gonzalez, winners of @Zoho 's "From the Ground Up" AI Hackathon. Their project, Dos Ojos, uses AI, satellite imagery, and drones to help farmers manage irrigation. #Zoho #AI
Microsoft Maia Chips to Become Primary Inference System
Satya Nadella announced Microsoft will make in-house Maia chips the primary inference system, scaling to gigawatts, reducing reliance on NVIDIA and replacing AMD chips.
TypeSafe Launches First System One Model Jev
TypeSafe AI released Jev, its first System One model, deciding 100x faster and cheaper than frontier LLMs. TypeSafe lists nine jobs where Jev engineering replaces pricey LLM calls.
Images 2.5 Brings Chinese Poetry Artistry to Life
A user creates Chinese-style illustrations with Images 2.5, capturing artistic conceptions from classical poetry: fine-line coloring, clouds and tree canopies, blank space, and poetic light-shadow effects.
CodeRabbit Claims to Halve Review Time and Bugs
CodeRabbit promotes its AI code review tool, claiming it can cut code review time and bugs in half, and offers a free trial signup.
Algorithm Discovers Routes from Elevation Data
A fun demonstration: an algorithm works from elevation data and discovers routes from first principles, showing data-driven path exploration.
Images 2.5 Creates Stunning Contour-Line Landscapes
A user experiments with Images 2.5 using one contour-line technique across four bold color schemes: glowing coral and deep blue, vermilion and cobalt, mineral flows, and neon terrain.
Google Antigravity Free for 10,000 Scientists
Google Antigravity is offering its highest usage tier free for six months to 10,000 scientists, with applications now open.
Parody Video: GPT-6 Passes the Dishwasher Benchmark
A parody ad for GPT-6 Astra circulates online, joking that GPT-6 achieves SOTA on the dishwasher benchmark, handling long tasks and stubborn stains alike.
Pancake gives coding agents customer-finding tools
Coding agents can now ship products end to end but lack customer-finding tools. Pancake is one MCP server giving Claude or Codex lead search across 40+ providers, enrichment, and sending from your own account.
Ajay's eval data tooling called promising entrant
Hamel Husain says he has known Ajay for a while and that Ajay is focused on tools for viewing data for evals, calling it a promising new entrant in the eval tools space.
kandev runs parallel AI coding tasks for one review
GitHubProjects highlights kandev: run many AI coding tasks in parallel, review all edits in one workspace, self-host without telemetry, and explore via osp.fyi/kandev.
Open-sourced phone brain trainer while agents work
modaleap jokes that employees are having too much fun with Agent, so he open-sourced a small phone brain-training gadget to use while waiting for agents to finish work.
AonicClub applies agents to physical goods sourcing
The post says AI rebuilt coding and legal drafting, yet supplement and grocery prices have not fallen. AonicClub applies agents to physical goods, handling sourcing and operations work.
harborai raises $10M seed for agent accountability
The post says the first agent question is who pays when it breaks, while most AI risk talk stops at hallucination. harborai, founded by four Stanford graduates, raised a $10M seed round.
XPENG G9L Paris debut set for October 12
XPENG teases the G9L European launch at the Paris Motor Show, framing it around Physical AI, a new intelligent class, and a live stream on October 12.
Google AI post shows only engagement metrics
The post contains only the Google AI handle and engagement counts, with no verifiable product, research, or timing details to extract.
CEOs want AI-savvy marketing leaders
Jason Lemkin says it is stressful to be a CMO or senior marketer; he knows 100+ CEOs desperate for an AI-pilled marketing leader, not someone running the 2024 playbook.
Meta AI chief explains small Muse team logic
Meta Chief AI Officer Alexander Wang says Meta knew Muse was the year's most important product yet kept a small focused team, because org bloat would break the coherence an agent needs for progressive user trust.
How to spot real videos versus AI videos
A guide to distinguishing real videos from AI deepfakes, emphasizing fact-checking, digital awareness, cybersecurity, and fraud risks on social media.
AI cybersecurity incidents may be iceberg tip
A brief warning suggests visible AI cybersecurity incidents may be only the tip of the iceberg, implying broader risks remain underestimated.
Aethir says 2.4B+ compute hours delivered
Aethir reports 2.4+ billion compute hours delivered globally, framing decentralized infrastructure as an operating network serving enterprise-scale production inference workloads.
Ex-Google CEO Schmidt: AI may hit cash wall first
Ex Google CEO Dr. Eric Schmidt argues AI may face a cash limit before an energy limit; at 6%, $1T debt costs $60B yearly, challenging perf-per-watt assumptions.
WSJ: AI compute scramble tops Silicon Valley
The post cites WSJ reporting that AI compute is now Silicon Valley's overriding priority, citing Google's $2.7B 2024 acquisition, reportedly to hire founder Noam.
RSI is systems engineering, not just models
The view says recursive self-improvement depends on the research agent's environment: callable resources, permissions to change things, and experiment loops mirroring production.
Adobe study: consumers trust AI search summaries
An Adobe study finds 63% trust AI search summaries at least as much as traditional results; inaccurate product details make 59% lose brand trust, risking immediate sales loss.
Barclays debates $1T AI capex by 2028
Brad Rogoff and Tom O'Malley debate a lab-driven AI capex forecast reaching $1T by 2028, nearly $300B above consensus, on The Flip Side.
3-year GPU commits can cost more than ownership
The poster says on-demand GPUs are scarce, so teams sign long offtakes; competitive rates require 3+ year commitments, while shorter terms raise hourly prices.
Open-source ICTRL lets agents control iPhone apps
The post highlights ICTRL as an open-source library letting computer or phone agents control iPhone apps, with the repo link in the comments.
A real Personal Agent workflow: one manual approval, that's it
The author breaks down a live X content automation: two bots split duties, run three daily tasks, and deliver videos to Xiaohongshu, Douyin, Shipinhao and Kuaishou — with human approval only once.
On-Site, a portfolio company, is hiring an AI engineer
Investor @iamazfer shares that one of his portfolio companies, On-Site, is hiring an AI engineer. He describes a strong team with a backlog of customers, and welcomes referrals of good candidates.
Screenweaver: screenplay to film in one AI pipeline
Screenweaver promotes its tool, which moves from screenplay to storyboard to finished film within a single AI pipeline, offering a free starting option.
Russia's Top Supercomputer Chervonenkis Reportedly Destroyed in Strike
Chervonenkis, Russia's most powerful supercomputer, was likely destroyed in a Ukrainian attack. It delivered 21-29 petaflops, prompting comparisons with the EU's top system, Germany's Jupiter.