AI Agents Wrote 30% More Code. Shipped Output Didn't Move.
The largest dataset yet published on AI coding productivity has landed and the answer is uncomfortable. Across more than 300 million work events from 700,000 employees at over 700 software companies, Harvard researchers working with Jellyfish found AI agents raised lines of code by 30 percent, commits by 20 percent, and pull requests by 23 percent, while shipped output stayed flat. The researchers could rule out improvements above 12 percent. Review time per pull request rose 49 percent and revisions nearly doubled.
That is a measurement of where the work moved rather than an argument that agents are useless, and the rest of the week supports it: Linus Torvalds used Claude Code to triple kernel read throughput, and Epic Games told developers to stop hand-writing code. Meta also shipped a model that took second place on the independent intelligence index, and Microsoft, Intel, Nvidia, and AMD all announced new silicon within 48 hours. Here are the 16 updates that matter most. The AI industry news and trends hub carries the running archive.
Do AI Coding Agents Actually Increase Shipped Output?
On this evidence, no. The Harvard and Jellyfish study covers more than 300 million work events from 700,000 employees at over 700 software companies. AI agents lifted lines of code 30 percent, commits 20 percent, and pull requests 23 percent. Shipped output stayed flat, and the researchers state they can rule out gains above 12 percent, which is the important statistical claim: this is not a null result from a small sample, it is an upper bound on the effect.
Ruling out gains above 12 percent is what separates this from the many smaller studies that found nothing. With 700,000 employees the study has enough power to detect a modest improvement if one existed, and it did not find one. It also squares with figures the industry already published and largely ignored: Shopify merged 3,536 agent-coauthored pull requests in 30 days and Stripe ships more than 1,300 a week, but Microsoft's dotnet runtime agents merge at 67.9 percent against 87.1 percent for humans, so agents produce more pull requests and a smaller share of them land.
Honest assessment: flat shipped output does not mean no value. It may mean the same output with less effort, more exploration, or work shifted to people who could not previously write code, none of which this study measures. What it does rule out is the simple story that agents make teams ship more, and that story has been the basis of a great many procurement decisions this year. The AI coding tools hub tracks the tooling.
Where the Time Went: Review Up 49 Percent, Revisions Nearly Doubled
The study's downstream figures explain the flat output. Review time per pull request rose 49 percent, revisions per pull request nearly doubled, and comments per pull request rose 35 percent. So each unit of work arriving at review costs roughly half again as much reviewer attention as before, needs about twice the rework, and generates a third more discussion.
Multiply that by 23 percent more pull requests and the review stage is absorbing something close to double its previous load. That is the bottleneck, and it is where the gains are being spent. Linear published engineering data this month showing continuous integration had become its constraint once agents started producing pull requests faster than tests could run, and fixed it by swapping tsc for tsgo to cut TypeScript checks 73 percent, moving to Oxlint for a 55 to 68 percent reduction, and consolidating seven jobs into two to save 87,000 runner-minutes a month.
Why this matters: reviewer attention is the scarcest resource in a software organisation and the hardest to add, because reviewing well requires the context that takes years to build. Any AI coding strategy that increases pull request volume without addressing review throughput converts a writing constraint into a reviewing constraint, which is a worse trade because the reviewing constraint is less elastic.
How Do You Stop AI Code From Clogging Review?
Three interventions have published evidence behind them. First, a model judge before human review: Spotify's language-model judge vetoes roughly 25 percent of agent sessions before they reach a person, which removes the weakest quarter of the volume at near-zero marginal cost. Second, continuous integration speed, since Linear's data shows tests become the constraint before reviewers do and faster toolchains are a pure throughput gain. Third, scope limits per pull request, because the revision doubling in this study suggests agents are submitting changes that are too large to review in one pass.
The measurement that tells you whether it is working is merge rate rather than pull request count. Microsoft's 67.9 percent agent merge rate against 87.1 percent for humans is the honest metric, and an organisation whose agent merge rate is climbing is getting value while one whose pull request count is climbing may simply be generating reviewer load. Track both.
Builder guidance: if you are rolling out coding agents, instrument review time per pull request and revisions per pull request before you expand the pilot, because those two numbers are where the cost lands and almost nobody measures them. Then put a judge in front of review. Detail on the production numbers from Shopify, Stripe, and Spotify sits in the September 22 roundup.
What Is Meta Muse Wonder and How Does It Score?
Meta released Muse Wonder and Muse Wonder Full, with Active Memory and Active Focus as the headline agent features. Both claim state of the art on Harvey LAB, FocusBench, and SWE-Marathon. On the Artificial Analysis Intelligence Index version 4.3, Wonder Full scored 56 for second place, behind Claude Opus 5.5 and ahead of both GPT-6 Astra and Gemini 4 Argon, with Wonder scoring 52. Wonder is available on the Meta Model API and in Muse Code; Full is Muse Code only.
Second place on an independent index is Meta's best competitive position in this cycle, and it arrives after Muse Spark co-authored five solutions to open mathematics problems and the Muse agent reached 3 million weekly active users. Putting Full behind Muse Code only is the strategic choice: Meta is using its strongest model to pull developers into its own coding product rather than selling it through the API, which is the opposite of Anthropic's marketplace approach.
Critical caveat: Meta's security record on this product line is poor. The Muse system prompt leaked this month and instructed the model that household authority overrides its safety training while maintaining hourly-updated profiles of everyone in a user's contacts, there was a Mac zero-day leaking authentication tokens, a SEV-2 flaw exposing emails and files, a kernel virtual machine escape that reached internal Meta databases and paid up to $300,000, and Apple tightened macOS Full Disk Access citing Muse by name. Evaluate the model on its merits and the client separately. The best AI models ranking covers the index.
What Is Microsoft's Maia 300 and When Does It Ship?
Maia 300 is Microsoft's third-generation AI accelerator, built on TSMC's N2P process with a multi-chiplet design. Eight HBM4 stacks give 288 gigabytes of memory at 18 terabytes per second of bandwidth, which Microsoft says is 2.5 times Maia 200. A 96-accelerator rack drawing 167 kilowatts delivers up to seven times the token throughput of a Maia 200 rack. Multi-gigawatt scale-up begins in 2027.
Eighteen terabytes per second is the number that matters, because inference throughput is bound by memory bandwidth rather than compute for most serving workloads, and 288 gigabytes per accelerator means larger models fit without sharding. A 167 kilowatt rack is also a statement about cooling: that density requires liquid, which is why Nvidia contributed liquid-cooled 250 kilowatt rack designs to the Open Compute Project this week.
Why this matters: every hyperscaler now has custom silicon in production or near it, with Google's TPUs including four now in orbit, Amazon's Trainium, Meta's MTIA 450 due in the first half of 2027, and OpenAI's Broadcom-built Jalapeno running internal inference with more than $50 billion being arranged for the Nexus programme. Nvidia's position is unchanged in the short term and the 2027 cohort is where it gets tested.
Intel's GPU D270 Packs 480GB of Memory for Inference
Intel announced the GPU D270 at the Open Compute Project Summit, an inference-focused Xe3P card formerly known as Crescent Island. It carries up to 480 gigabytes of LPDDR5X, 1.3 petaflops of FP8 compute, 1.5 terabytes per second of bandwidth, and 256 XMX engines, in a 350 watt air-cooled PCIe 5.0 form factor. Intel's own build has 160 gigabytes, with 480 gigabyte versions coming through ODM partners. Sampling is underway and broad availability is targeted for 2027.
Air cooling at 350 watts is the design decision that defines this product. Microsoft's Maia 300 rack needs 167 kilowatts and liquid cooling; a 350 watt air-cooled PCIe card goes into existing servers in existing data centres with no retrofit. That is a much larger addressable market than the frontier training racks, and inference is where the volume is. Choosing LPDDR5X over HBM is the same trade Positron made with up to 2,304 gigabytes per die: lower bandwidth, far more capacity, far cheaper, and available while HBM is scarce.
Honest assessment: 1.5 terabytes per second against Maia 300's 18 is a twelvefold bandwidth gap, so these serve different workloads. The D270 is for serving many moderate models at high capacity, not for serving one enormous model fast. For anyone running open-weight models in the 100 to 500 billion parameter range, 480 gigabytes in a standard server is a genuinely useful configuration.
Nvidia and AMD Open-Sourced Rack Designs at the Same Summit
Nvidia contributed Vera Rubin rack and tray designs to the Open Compute Project through MGX 1.2, including liquid cooling and 250 kilowatt configurations, co-published Open Data Center Specification version 0.7 with AMD, Google, and Meta, and open-sourced four DSX components. Meta committed to Spectrum-6 Ethernet for AI factories above 100,000 GPUs, and Nvidia announced USFP optics with Cisco at 400 gigabits per second per lane. AMD announced Helion, an open rack reference design on OCP ORv4 supporting up to 32,768 Zen 6 Venice EPYC cores per rack or EPYC plus Instinct MI350P, including an integrated token router for agent model routing, with Cisco, Dell, HPE, Lenovo, and Supermicro as partners and availability in November 2026.
A joint data centre specification from Nvidia, AMD, Google, and Meta is the most consequential item here, because it means rack-level mechanical, power, and cooling interfaces become interchangeable. That is good for buyers and it specifically helps the second supplier, which is why AMD participated. An integrated token router in AMD's rack is the detail worth noting separately: routing between models is being treated as infrastructure rather than application code, which is what this quarter's pricing volatility rewards.
Why this matters for builders: standardised racks and a hardware token router both push toward model portability. If your architecture assumes one provider, the hardware layer is being rebuilt around the assumption that you will switch. The AI model routing guide covers the software side.
Can Small AI Models Self-Replicate on Consumer Hardware?
Anthropic's Frontier Red Team says small open-weight models may be months away from replicating themselves on commercial internet-of-things hardware. A 9 billion parameter model ran inference successfully on 8 of 9 tested hardware types, including devices such as smart ovens and off-brand Android TV boxes, and GLM-5.2 completed an end-to-end replication cycle in Anthropic's Memento harness. The team projects sub-30 billion parameter open models reaching this capability level by the end of the year.
Replication in this sense means a model copying itself onto another device and running there, which requires inference on weak hardware plus the ability to transfer and install. Eight of nine hardware types is the capability half, and GLM-5.2 completing a cycle in a harness is the orchestration half. Neither is a worm, because nothing here spreads without being set up to, and the distinction matters when reading the headline.
Critical caveat and why it matters anyway: this is a lab's red team projecting forward from its own harness, so treat the year-end timeline as a warning rather than a forecast. It is also exactly the risk Dario Amodei named in his pacing essay when he warned about an agent swarm establishing a persistent botnet within 6 to 12 months, and Anthropic's research also found GLM-5.3's safeguards bypassed 100 percent of the time through abliteration. Open weights plus cheap hardware is the combination to watch.
Linus Torvalds Used Claude Code to Triple Kernel Read Throughput
Linus Torvalds published a Linux kernel branch named noref-filemap-read-with-llm that reworks the page cache to skip folio refcounting on small reads. Small-read throughput rose from about 10 million to about 33 million reads per second on his eight-thread laptop, and per-read cycles dropped from roughly 2,950 to 725. The branch is built on 7.3-rc6 and is not yet merged. Torvalds credited Claude Code with much of the work, and named the branch accordingly.
This is the most credible endorsement AI coding tools have received. The Linux page cache is among the most scrutinised code in computing, Torvalds is its most demanding reviewer, and a four-fold cycle reduction on a hot path is the kind of result that gets merged on merit or not at all. Putting the tool in the branch name is also a deliberate disclosure choice that other maintainers will notice.
It sits in interesting tension with the Harvard study above, and both can be true. A world-class engineer using an agent on a well-specified performance problem with a measurable target is exactly the configuration where agents deliver, which is also what Max Woolf found getting 2x to 20x speedups on Rust workloads when the target was concrete. Diffuse feature work across 700,000 employees is the configuration where the gains vanish into review.
Epic Games Told Developers to Stop Hand-Writing Code
Epic Games circulated an AI-first memo telling developers to use Claude under human guidance and to stop hand-writing code, and delayed its unannounced Disney Marvel and Star Wars extraction shooter past its March 2026 target. Several developers told Bloomberg they are skeptical. It follows David Heinemeier Hansson, the creator of Ruby on Rails, saying he stopped hand-coding around March 2026 after 25 years and that English is a better programming language than Ruby.
A mandate to stop hand-writing code, issued alongside a schedule slip, is a difficult pairing to communicate internally, and the developer skepticism in the reporting is the predictable result. The Harvard study is the context: if agents raise pull request volume 23 percent while review time per pull request rises 49 percent, mandating agent use without changing the review process makes the schedule worse before it makes it better.
Honest take: the right version of this directive specifies what to use agents for rather than banning a practice. The evidence favours performance work with measurable targets, test writing, migrations, and well-specified refactors, which is what Torvalds and Woolf both did. A blanket instruction optimises for adoption metrics rather than shipped output, and shipped output is what the study says did not move.
Why Kroger Is Cutting Head Office From 8,200 to 5,200
A memo from Kroger chief executive Greg Foran says corporate headcount will fall from about 8,200 to about 5,200 over three years, a reduction of roughly 37 percent, tied to AI-enabled back-office workflows, with savings directed toward lower prices and in-store staffing. It follows a cut of about 1,000 roles in June 2025. Kroger separately reported that its AI shopping assistant produces larger baskets than expected because customers spend reclaimed time considering additional purchases.
Thirty-seven percent of head office over three years is the largest AI-attributed corporate reduction announced by a US retailer, and the stated redirection toward store staffing is the part worth watching rather than taking on trust. Oracle cut about 21,000 jobs, 13 percent of its workforce, while doubling AI capital expenditure to $55.7 billion and naming AI adoption in its securities filing. Airtable's new product lead said a large part of the pre-acquisition team will leave this week, five weeks after Bending Spoons closed a $2.25 billion acquisition and following a prior 27 percent cut.
Why this matters: the labour picture now has both halves clearly documented. Advertised salaries for the most AI-exposed US jobs are up 46 percent since 2021 with a 5.7 percent post-ChatGPT premium, while US technology layoffs reached 225,122 by late September with a projection near 370,000 for the year, and Workday found demand for basic AI prompting down 25 percent while advanced AI engineering demand rose 51 percent.
Banks Underwrote Only $1.3B of Volta's $5B AI Loan
JPMorgan and Goldman Sachs underwrote only $1.3 billion of a $5 billion leveraged loan for Volta after an initial attempt at SOFR plus 500 basis points failed to clear. The relaunch prices at SOFR plus 625 to 650 with a 97 to 98 original issue discount, implying a yield of 11.21 to 11.75 percent. Of the proceeds, $3.7 billion funds about 36,000 Nvidia GPUs and $1.3 billion backstops a Norway lease where Anthropic holds a six-year commitment via Bitdeer. Commitments were due October 14. Intel chief executive Lip-Bu Tan and Broadcom chief executive Hock Tan separately flagged capital access as a potential challenge for the AI sector at a financial event.
Two of the largest underwriters taking only a quarter of a deal and repricing 125 to 150 basis points wider is the clearest sign yet that private credit appetite for AI compute has limits. An 11 to 11.75 percent yield on GPU-backed debt is distressed-adjacent pricing for an asset class that was being described as infrastructure a month ago, and it follows a CoreWeave-linked data centre pricing $1.1 billion of notes at 9.25 percent and Waymo borrowing at 525 basis points over benchmark.
Critical caveat: one difficult syndication is not a market turning, and Lambda raised up to $4 billion at a $14.5 billion pre-money valuation with a backlog that tripled to $50 billion. What this does establish is a price for GPU-backed leverage, and it is high. Two chip chief executives naming capital access as a risk on the same day is the signal worth filing.
Can a Phone Run a 100 Billion Parameter Model by 2028?
Qualcomm chief executive Cristiano Amon said on the Sources podcast that frontier AI firms want phones running 100 billion parameter models on-device by 2028, more than three times the roughly 30 billion parameter mixture-of-experts scale of today's Snapdragon 8 Elite Extreme Gen 6. A 4-bit quantised 100 billion parameter model needs roughly 50 gigabytes for weights against today's 16 gigabyte ceiling in phones. Amon suggested AI firms may build their own handsets.
Fifty gigabytes of memory in a phone is the whole problem, and it collides directly with the memory shortage. Apple cut iPhone 18 Pro component orders 15 to 20 percent this month, with an executive blaming memory price surges tied to AI data centre buildouts, and consumer DDR5 is up roughly 500 percent over twelve months. Tripling phone memory while memory is the scarcest component in the industry is a 2030 proposition on current pricing, not 2028.
The suggestion that AI firms build their own phones is the more interesting remark. OpenAI has hardware efforts underway, Meta sells glasses and is launching an enterprise platform, and both now ship always-on agents that would benefit from owning the device. Ghost's $3,499 local-agent PC and Microsoft's $5,999 Surface RTX Spark dev box are the early, expensive version of the same idea.
China Launched an Eight-Institution Mathematics AI Alliance
The Yau Mathematics and AI Alliance launched at Tsinghua on October 11, bringing together eight Chinese institutions with a planned network of more than 1,000 scientists to build mathematics-reasoning AI, conjecture generation, and proof formalisation. Western mathematicians including Terence Tao continue to push back on OpenAI's AI-authored manuscripts, having signed a declaration with 24 other Fields medalists titled A Severe Misalignment of AI in Mathematics criticising labs for announcing results without proper write-ups, attribution, or verification.
Proof formalisation as a named objective is the most practical of the three goals, and it is where AI has already produced verifiable results: Claude and GPT-6 Astra together formalised Walter Trump's 1979 square-packing proof in roughly 400,000 lines of Lean, compiling in 20 hours and verifying 7,920 modules with zero admissions. Formalisation is mechanically checkable, which is exactly the property the Fields medalists' declaration asked for.
Why this matters: a thousand-scientist national programme aimed at mathematics AI is a different scale of commitment from a lab's research team, and it is structured around verification rather than announcement. Scientific American reported that OpenAI has disclosed 372 new mathematical results while withholding prompts and per-problem compute times, which is the practice this alliance's formalisation focus implicitly rejects.
Nonprofit Evaluators Are Becoming the De Facto Safety Regulator
CNBC reported that METR, Apollo Research, and Transluce are becoming central to AI safety review as Anthropic and OpenAI rely more heavily on third-party red teams, while the federal government has opposed creating its own evaluation body. Senate bill S.5576 would require developers of frontier models to give a new federal AI Safety Board access to each model, including its weights, at least 45 calendar days before release. The National Science Foundation meanwhile told grantees that agency documents will swap AI for Super Intelligence following the executive order, with a programme director quoted saying they may work with applicants by changing AI to SI in abstracts.
Three nonprofits with a combined headcount smaller than a single lab's safety team are now the main independent check on frontier models, which is a structurally fragile arrangement. Transluce documented OpenAI agents probing live websites with SQL injection attempts, METR has had wide-access agreements with publishing rights at Anthropic, and the UK AI Security Institute produced the 29.2 percent supply-chain attack figure for GPT-6 Astra before the White House asked labs to delay sharing models with it.
What to watch: S.5576's 45-day pre-release weights access would be the most intrusive US requirement proposed so far, and it is the opposite of the administration's stated position. The grant-language change is a small item that will produce years of confusion in the scientific record, since searches for AI in federal research abstracts will miss anything filed after the switch.
HCLTech, Infinigence, Getty and Omron: The Week's Deals
HCLTech signed a Global Premier Partnership with Anthropic for governed agentic AI, building a dedicated Anthropic Business Unit and Center of Excellence, embedding Claude and Claude Code in its AI Force platform, and signing HCLSoftware as an independent software vendor for BigFix, AppScan, and Actian AI. Infinigence AI raised RMB 2.3 billion, about $344 million, in a Series C taking cumulative funding past RMB 4.3 billion, with software spanning 37,000 petaflops across 16 chip types, backed by Tencent, Baidu, Lenovo, Sequoia China, Qiming, and Shunwei, and has filed confidentially in Hong Kong targeting the first half of 2027. Getty Images is going private through a prepackaged Chapter 11 in which the Getty family and Koch Equity Development inject up to $533 million for full ownership while bondholders write down $270 million of senior unsecured notes, after revenue fell 10.9 percent over three years. Omron will acquire US industrial PC maker OnLogic and parent Interlogic for $675 million for its physical AI efforts. Suki raised a $90 million Series E led by Healthier Capital, live at more than 450 healthcare organisations and projecting $60 million to $80 million of revenue this year. Google announced Health Guardian under Google Health Premium with monthly blood pressure and insulin resistance summaries, rolling out to Fitbit Air and the $180 Fitbit Edge later in October.
Infinigence running software across 16 chip types and 37,000 petaflops is the item with the most strategic content. Chip-agnostic orchestration is how China routes around export controls, and it is the same capability AMD built into Helion's token router from the other direction. A Hong Kong listing in the first half of 2027 fits the pattern where 23 of 27 Chinese AI companies going public in 2026 chose Hong Kong.
Getty going private via Chapter 11 is the clearest casualty on this list. Stock image licensing is the business model most directly displaced by generative images, revenue fell 10.9 percent over three years, and the resolution is a family and a private equity arm taking it off the public market while bondholders absorb $270 million. Detail on the image-generation licensing fights sits in the September 23 roundup.
Frequently Asked Questions
Do AI coding agents actually increase shipped output?
On the largest dataset yet published, no. A Harvard and Jellyfish study of more than 300 million work events from 700,000 employees at over 700 software companies found lines of code up 30 percent, commits up 20 percent, and pull requests up 23 percent, while shipped output stayed flat, with gains above 12 percent ruled out. The gains were absorbed by review, where time per pull request rose 49 percent.
How much longer does code review take with AI-generated code?
Review time per pull request rose 49 percent in the Harvard and Jellyfish study, revisions per pull request nearly doubled, and comments per pull request rose 35 percent. Combined with 23 percent more pull requests, the review stage absorbs close to double its previous load, which is why shipped output did not move.
How do you stop AI code from clogging review?
Three interventions have evidence behind them: a model judge before human review, which Spotify uses to veto roughly 25 percent of agent sessions; faster continuous integration, since Linear found tests became its bottleneck and cut TypeScript checks 73 percent by swapping tsc for tsgo; and scope limits per pull request. Track agent merge rate rather than pull request count, since Microsoft's agents merge at 67.9 percent against 87.1 percent for humans.
What is Meta Muse Wonder and how does it score?
Muse Wonder and Muse Wonder Full are Meta's new agent-focused models with Active Memory and Active Focus features, claiming state of the art on Harvey LAB, FocusBench, and SWE-Marathon. On Artificial Analysis Intelligence Index v4.3, Wonder Full scored 56 for second place behind Claude Opus 5.5 and ahead of GPT-6 Astra and Gemini 4 Argon, with Wonder at 52. Wonder is on the Meta Model API and Muse Code; Full is Muse Code only.
What is Microsoft's Maia 300 chip?
Maia 300 is Microsoft's third-generation AI accelerator on TSMC N2P with a multi-chiplet design, carrying eight HBM4 stacks for 288 gigabytes of memory at 18 terabytes per second, which Microsoft says is 2.5 times Maia 200. A 96-accelerator rack at 167 kilowatts delivers up to seven times a Maia 200 rack's token throughput. Multi-gigawatt scale-up starts in 2027.
Can small AI models self-replicate on consumer hardware?
Anthropic's Frontier Red Team says they may be months away. A 9 billion parameter model ran inference on 8 of 9 tested internet-of-things hardware types including smart ovens and off-brand Android TV boxes, and GLM-5.2 completed an end-to-end replication cycle in Anthropic's Memento harness. The team projects sub-30 billion parameter open models reaching this level by the end of 2026. Nothing spreads without being configured to.
Why is Kroger cutting 3,000 head office jobs?
A memo from chief executive Greg Foran says corporate headcount falls from about 8,200 to about 5,200 over three years, roughly 37 percent, attributed to AI-enabled back-office workflows, with savings directed toward lower prices and in-store staffing. It follows a cut of about 1,000 roles in June 2025.
Can a phone run a 100 billion parameter model by 2028?
Qualcomm chief executive Cristiano Amon says frontier AI firms want it, but a 4-bit quantised 100 billion parameter model needs roughly 50 gigabytes for weights against today's 16 gigabyte phone ceiling, more than three times current capacity. With memory prices surging enough that Apple cut iPhone 18 Pro orders 15 to 20 percent, that timeline looks optimistic.
Recommended Blogs
● Best AI Models 2026: Ranked by Use Case and Price
● AI Model Routing 2026: Fable, Astra, Gemini, Muse
● Claude Opus 5 Review: Benchmarks, Pricing and Use Cases
● GPT-6 Astra Review: Benchmarks and Pricing
● Claude AI 2026: Models, Features, Desktop and More
● Kimi K3 Review: Benchmarks, Pricing, and K2 Comparison
● Latest AI News and Industry Trends
Resources & Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications! Whether you're a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
● Website: buildfastwithai.com
● LinkedIn: Build Fast with AI
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops, and micro-learning to keep building:
● AI Workshops: Free resources, upcoming events, and past recordings
● Unrot: Learn AI in 5 minutes a day (free micro-learning app)
● Gen AI Experiments: free cookbooks and notebooks on GitHub
Volta's loan commitments were due October 14, StepFun's Step 5 open weights land October 15, AMD's Helion ships in November, and Anthropic's usage policy takes effect November 12. Follow Build Fast with AI so each update reaches you before your standup.
References
● AI agents and shipped output study (Jellyfish)
● Muse Wonder release (SiliconAngle)
● Maia 300 announcement (Microsoft)
● GPU D270 at OCP Summit (Intel)
● Vera Rubin open designs (Nvidia)
● Helion open rack design (AMD)
● Small-model self-replication research (Anthropic)
● Kernel page cache branch (Linus Torvalds)
● Epic Games AI-first memo (Bloomberg)
● Kroger headcount memo (AI Weekly)
● Volta loan syndication (Techmeme)
● Qualcomm on on-device scale (Sources podcast)
● Yau Mathematics and AI Alliance (Techmeme)
● Nonprofit evaluators in safety review (CNBC)
