Highlights
What happened at the companies we cover, by day, newest first. Every item links to its sources.
Showing 1–20 of 25 highlights
Friday, Oct 23 25 stories
Fireworks grows staff fast and courts builders with a Microsoft and NVIDIA buildathon
HiringX
A workforce tracker reported Fireworks grew headcount by more than half in 90 days to about 448 people, ranking it first among fast-growing AI companies. Separately, Fireworks announced its Deploy buildathon with Microsoft and NVIDIA for Nov 2 in San Francisco, challenging teams to show tuned open models beating a closed frontier model on a chosen eval. The moves fit recent reports of a much larger office lease in San Mateo and rumored funding talks. Engagement on these posts was modest.
Sources
Hugging Face publishes a lookup of which Hub models run unmodified on AWS Trainium
LaunchNews
Hugging Face launched a Space that reports whether popular Hub models run without changes on AWS Trainium through the TorchNeuron PyTorch backend. Each entry lists the result, how many Trainium cores the model was sharded across, wall time including compilation and peak device memory, plus a script to reproduce the run. Views follow the Hub's trending and download rankings, and models left out are flagged with a reason. The results are also published as a dataset, though only Transformers text and image-text models are covered so far.
Sources
Modal shuts down its hosted Notebooks product
NewsOther
Modal retired Modal Notebooks effective October 23, saying users now prefer working from their own IDE or agent harness. Existing notebooks remain downloadable as standard IPython files with no deadline. The move trims a hosted interactive surface in favor of code-first and agent-driven workflows. It fits Modal's recent focus on sandboxes, training and inference primitives rather than a notebook UI.
Sources
Together AI says Factory tested Reflection's Beam model through it before launch
DiscussionX
Together AI's official account said Factory got early access to Reflection AI's Beam model through Together ahead of Beam's launch. According to the post, Factory used that window to test Beam on real coding and agentic workflows, including tool-error recovery, conflict handling and final repo verification. Together is presenting itself as the pre-launch evaluation channel between open-model labs and agent builders. An SGLang contributor reposted it as an example of the open AI ecosystem working together. Engagement was modest, at about 21 likes on the original post.
Sources
Nvidia and TCV back Higgsfield in Series B extension as it passes $1B run rate
NewsNewsBlueskyX
Higgsfield announced that TCV and Nvidia invested in an extension of the Series B it raised in August, led by DST Global. It did not disclose the size of the extension or the valuation. The company said it crossed $1 billion in annualized revenue in September, about 18 months after launch. It reported more than 40 million users, with enterprise as its fastest-growing segment. Bloomberg reported this is Nvidia's first investment in Higgsfield, and the announcement coincided with Jensen Huang's first visit to Kazakhstan, where the company has deep roots.
Sources
Crusoe's Itaipu-powered Paraguay AI cloud draws questions over power contract and Brazil's share
NewsNewsXBlueskyHN
Crusoe announced a partnership with Paraguayan developer Araico to build AI cloud capacity next to the Itaipu dam substation. The first phase is due in mid-2027, with a path to 106MW of IT capacity and options for about 280MW. Regional coverage noted that the US State Department fact sheet gave no tariff, amount or power-utility details. A Paraguayan outlet reported, citing Araico and a utility union, a long-term supply deal starting near $45/MWh. On X, Brazilian commentators debated whether redirecting Paraguay's Itaipu surplus to the project hurts Brazil.
Sources
Modal details kernel-level serving work for Sakana AI's ~1T-parameter Namazu API
NewsX
Modal described how it optimized Sakana AI's Namazu API, which runs a roughly 1-trillion-parameter model on its serverless GPU pool. The work used selective NVFP4 quantization on the routed experts while keeping attention and shared layers in BF16, plus a DFlash draft model for speculative decoding that Modal says raised acceptance by about 25%. Modal also cited in-house prefix KV-cache routing and serverless autoscaling. Researchers shared the case study as an example of serving a frontier-scale mixture-of-experts model on four Blackwell GPUs.
Sources
Rival CEO claims Harvey discounts buyers who skip trialing competitors
DiscussionX
Spellbook's CEO posted that Harvey is offering lower prices to customers who agree not to trial a competitor, calling the practice not customer-friendly. His evidence is a screenshot of an unattributed message describing tiered pricing tied to whether, and when, a buyer runs competing trials. The screenshot does not clearly name Harvey, and Harvey has not commented, so this remains an unverified claim from a direct competitor. The post drew more than 100 likes, and replies framed it as defensive pricing amid Harvey's rivalry with Legora.
Sources
Cognition ships refreshed Devin Desktop app with a new agent command center
LaunchX
Cognition introduced a refreshed Devin Desktop app built around an agent command center that it says is smoother and more polished. Users say the release brings the desktop agent window in line with the web app, with shared session lists, automations, review, wiki and a command palette. The announcement drew strong engagement. Early users praised its stability with many parallel agents, and one said the older app crashed often enough to push them back to the CLI.
Sources
Robinhood Ventures discloses ~$10M stake in Cognition as Devin passes $1B run rate
NewsNewsX
Robinhood Ventures Fund I, a publicly traded closed-end fund, said it closed about $10 million of investments in Cognition. According to Fortune, the stake was part of Cognition's September Series E, which raised more than $2 billion at a $48 billion valuation; Robinhood is only now disclosing its participation. Robinhood's release says Cognition recently crossed $1 billion in annualized revenue run rate and lists Goldman Sachs, Citi and Nubank as Devin customers. Robinhood had already invested in OpenAI, and its fund head said AI coding is not a winner-take-all market.
Sources
Mercor CEO re-ups $100K data-licensing referral offer, citing 500+ partner companies
DiscussionX
Mercor's CEO again promoted an offer of up to $100,000 for each referred company that licenses its data to Mercor. He said Mercor has already partnered with more than 500 companies and is targeting firms with 100+ employees and data in code, gaming, manufacturing, healthcare or professional services. The post drew several hundred likes and many bookmarks. Some replies pushed back: one person who said they work in the field claimed typical referral payouts are closer to $10,000, and another argued that 500 partners is a small base for a company valued so highly.
Sources
Job-post counts show Clay becoming a default requirement for GTM engineer roles
DiscussionX
Several posters counted how many GTM engineer job listings name Clay as a required tool. One said 49 of about 60 posts they read this week mentioned Clay. Another found 28 of 43 sampled GTM engineer listings did, and noted that Clay itself popularized the 'GTM engineer' title through a newsletter, a hiring guide, a job board and a bootcamp. A minority view questioned paying for Clay's flexibility when a simple list is all that's needed.
Sources
Clay teams up with SpaceXAI to run Grok Bot agents inside Clay workflows
NewsX
Clay's official account said it is working with SpaceXAI to put Grok Bot to work inside Clay. Users will describe their starting data, the steps needed and the desired output in natural language, and the agent builds the Clay workflow. Clay said a live session on October 29 will show the agent researching companies, enriching contacts and drafting emails for reps to review. The agent will also fill in web forms, log into apps and pull Clay data from a phone.
Sources
Replit's forced security republish and two Oct 23 incidents frustrate some users
ComplaintXBluesky
Replit said it patched infrastructure, updated language tooling and blocked exploits after recent third-party software vulnerabilities, and asked some users to republish their apps. Several users replied that republishing failed or broke their sites, and that support had been slow to resolve tickets. Separately, Replit's status page logged two incidents on October 23: users couldn't access privately deployed apps for about 40 minutes, and HubSpot connectors stopped working later in the day. Replit Support said engineers were investigating individual cases.
Sources
1 more source
Chroma agrees to join Replit; Chroma Cloud to wind down over the next 60 days
NewsHNBlueskyX
Chroma, the company behind the popular open-source vector database ChromaDB, said it has signed a definitive agreement to join Replit. The combined goal is to build a 'context layer' so agents can draw on everything an organization knows, and Chroma's Foundation memory product will continue inside Replit. Replit's CEO and president were early investors in Chroma. Chroma Cloud will be wound down over the next 60 days, while open-source ChromaDB continues. Early reactions included questions about migration paths for Chroma Cloud users.
Sources
Arena teases an incoming Kimi model as 'Kibitz' is widely tied to Kimi K3.1
DiscussionXBlueskyNews
Arena's official account posted that a new Kimi model is coming and told followers to watch where it lands on the leaderboard, and the post drew over 2,000 likes. Moonshot separately posted a short teaser video, and community posts tied the anonymous Arena model 'Kibitz' to Kimi K3.1 and 'Kalmia' to Qwen 4, though neither tie is officially confirmed. Testers also reported new anonymous OpenAI image models on Arena and the Gemini 4 Carbon checkpoint. Some skeptics noted that codename sightings bring weeks of vibe-based claims before controlled comparisons are possible.
Sources
Inc. profiles why Gamma's CEO rebuilt the product after passing 100 million users
NewsNews
An Inc. column looks at why Gamma CEO Grant Lee chose to rebuild the product from the ground up even though it had passed 100 million users and $100 million in annual recurring revenue. Lee said the guardrails on margins, spacing and fonts that once made decks look polished had become limits as AI improved, and users needed more expressiveness. The rebuild shipped as Gamma 5 on October 6, adding wider visual styles, a freeform canvas alongside agent editing, enterprise connectors and design-preserving import and export. The column presents it as a response to AI output that increasingly all looks the same.
Sources
BFL says FLUX 3 Image Klein now ranks #1 among open-weight text-to-image models
NewsXBluesky
Black Forest Labs said FLUX 3 Image Klein ranks first in text-to-image among open-weight models on third-party benchmarks, and that its own testing shows the model is strongest at precise image editing. Posts tracking the Artificial Analysis leaderboard noted the 5.6B-parameter model leads Qwen-Image 2.1 by a single point. Some users welcomed the result and praised its hand rendering, while others were tired of rotating 'number one' claims and asked BFL to release open weights for FLUX 3 Dev editing instead. The model, released October 15, can be fine-tuned with LoRA and runs on consumer GPUs.
Sources
Baseten becomes inference partner for Mirendil's new base model
NewsX
Baseten said it will serve Mirendil's new base model as its inference partner. The company reported that in its own tests on private codebases the model had never seen, Mirendil-Base-Large scored the lowest bits per byte, a tokenizer-independent measure of perplexity, of any model it tested. The post drew a pointed reply from a well-known ML researcher asking whether Baseten's stack relies on open-source engines like vLLM, SGLang or TensorRT-LLM. The exchange got solid engagement, though Baseten's benchmark claims are self-reported.
Sources
Baseten launches an Environment Network of vetted partners to build custom RL environments
LaunchX
Baseten introduced the Baseten Environment Network, a shortlist of audited RL environment companies that will build custom post-training environments from each customer's own data. Partners such as Snorkel AI, Abundant, Proximal, Eigent, micro1, Preference Model, Arcanum and Taste Labs work directly with customers on environments, tasks and rubrics, while Baseten keeps the training and production inference. Baseten says it has done this work in-house for customers like Harvey, Lovable and Gamma and needed more expert teams to scale it. It also plans to open-source a set of frontier-quality environments along with its auditing and training methods. Several partners announced their participation on X the same day.
Sources