Apple Took Years to Catch Up. Kilo Code Took 6 Weeks--and It's Coming for Lovable, Cursor, Replit

TL;DR

XAI’s $20 billion upsized Series E (about $230B valuation) signals that only a few AI labs have the capital depth to endure multi-year scaling costs.

Briefing Cornell Notes

Briefing

XAI’s $20 billion upsized Series E—valued around $230 billion—lands at the center of a widening divide in AI funding: only a handful of labs now have the runway to survive the long, expensive scaling race. The money is earmarked for expanding XAI’s Colossus supercomputers in Memphis, where the company says it ended 2025 with more than 1 million H100 GPU equivalents across Colossus 1 and 2, and that Grok 5 is currently in training. XAI also claims roughly 600 million monthly active users across X and Grok apps, positioning it as the largest consumer-facing AI deployment outside Google and OpenAI.

The raise arrives amid a safety and regulatory backlash. Grok-generated inappropriate deepfakes of real people—including minors—triggering probes across the EU, UK, India, Malaysia, and France. Yet XAI still secured a Department of Defense deal, with Grok now serving as the DoD’s AI agents platform, and Grok also powers prediction markets including Poly Market and Call. The implication is that investors are treating early product and safety failures as solvable growing pains—backing companies that can keep scaling even while investigations continue. The funding landscape, as framed here, leaves OpenAI, Anthropic, and XAI with clear multi-year survival odds, while others face shorter timelines and more fundraising risk.

That financial reality is running in parallel with a high-stakes debate about how fast “AGI” could arrive. At Davos, Anthropic’s Dario Amodei and Google DeepMind’s Demis Hassabis agreed AGI is coming, but diverged on timing and the mechanics of progress. Amodei leaned toward AGI emerging in 2026 or 2027, driven by an accelerating feedback loop where AI writes code and humans review it; he also noted Anthropic engineers rarely write code by hand anymore. Hassabis was more conservative, putting the probability of AGI at 50% by the end of the decade, arguing that jobs aren’t easily automated because humans still supply the hard-to-replicate “last 5%” of skills. Both pointed to technical gaps that matter beyond employment: memory, continuous learning, and long-term reasoning—areas where today’s models still struggle, including a “memory wall” and reasoning that doesn’t yet match human-like long-horizon problem solving.

Meanwhile, Apple’s latest AI partnership signals a major shift in distribution power. Apple and Google announced a multi-year collaboration in which Apple’s next generation of foundation models will be based on Google’s Gemini and Google cloud technology. The deal reportedly costs Apple about $1 billion per year, and Google is said to be building a custom 1.2 trillion-parameter Gemini model for Apple—far beyond what Apple’s current models can deliver. The knock-on effect: pressure rises on OpenAI’s Sam Altman and Jony Ive to secure a third device strategy that can preserve OpenAI’s distribution footprint.

On the research front, DeepSeek published EnGram, a conditional memory architecture aimed at fixing a transformer weakness: lack of native knowledge lookup. By using short token sequences, hash-based retrieval from a large embedding table, and gating to filter results against context, EnGram reduces the need for expensive “reasoning tokens” and improves token efficiency—an approach framed as a step toward more factual memory. Finally, Kilo Code launched an app builder after a six-week sprint, targeting engineers with a VS Code-like, open-source-friendly platform strategy aimed at competing with Lovable, Replit, and Cursor. The central question becomes whether an engineer-first workflow can carve out space as “vibe coding” matures from novelty into reliable tooling.

Cornell Notes

XAI’s $20 billion upsized Series E (about $230B valuation) underscores that only a few AI labs—OpenAI, Anthropic, and XAI—now have the multi-year funding runway to win the scaling race. The raise happens despite Grok deepfake incidents and regulatory probes across multiple countries, yet XAI still landed a U.S. Department of Defense deal and expanded Grok’s ecosystem. At Davos, Anthropic’s Dario Amodei predicted AGI in 2026–2027 via AI-assisted coding loops, while Google DeepMind’s Demis Hassabis argued for a 50% chance by decade’s end and emphasized that jobs may be disrupted mainly at the “last 5%” of human skills. Apple’s Gemini-based model collaboration shifts foundation-model leverage toward Google’s distribution. Research progress like DeepSeek’s EnGram targets token-efficient “factual memory” by adding conditional retrieval to transformers.

Why does XAI’s $20B funding round matter even with ongoing safety investigations?

The round is framed as a signal that investors are prioritizing long-term scaling capacity over short-term product imperfections. XAI earmarks the capital for expanding Colossus supercomputers in Memphis, where it reports over 1 million H100 GPU equivalents across Colossus 1 and 2 and that Grok 5 is in training. Even after Grok deepfakes of real people—including minors—triggered probes in the EU, UK, India, Malaysia, and France, XAI still secured a Department of Defense deal, with Grok positioned as the DoD’s AI agents platform. That combination—continued scaling plus high-value contracts—helps explain why the funding timeline didn’t stall.

What disagreement about AGI emerged at Davos, and how does it connect to jobs?

Dario Amodei (Anthropic) leaned toward AGI appearing in 2026 or 2027, citing accelerating feedback loops where AI writes code and humans review it; he also said Anthropic engineers rarely code by hand anymore. Demis Hassabis (Google DeepMind) was more conservative, estimating a 50% probability of AGI by the end of the decade. On employment, Hassabis argued automation is limited because jobs aren’t easy to fully replace—if AI gets 95% of skills, it mainly increases the value of the remaining 5% humans do well. The transcript links this to mixed real-world employment signals: junior roles may be harder to get, but aggregate layoff impacts aren’t clearly tied to AI across the economy.

Which technical gaps were highlighted as central to whether AGI arrives on schedule?

Three areas were emphasized: memory, continuous learning, and long-term reasoning. Current models face a “memory wall,” meaning they don’t learn after release, and long-term reasoning still doesn’t match human-style problem solving. The transcript suggests Hassabis expects progress here, but not necessarily on the same timeline as Amodei’s 2026–2027 estimate. The practical takeaway: the next year or two may reveal which approach is closer to reality.

How does Apple’s Gemini-based model collaboration change the competitive landscape?

Apple and Google agreed that Apple’s next generation of foundation models will be based on Google’s Gemini and Google cloud technology. The deal is reported to cost Apple about $1 billion per year, and Google is said to be building a custom 1.2 trillion-parameter Gemini model for Apple. The transcript interprets this as a major loss for OpenAI because it shifts default AI experiences toward Gemini across platforms, not just Android. That increases pressure on OpenAI’s leadership—Sam Altman and Jony Ive are named—to deliver a third-device distribution strategy.

What is EnGram, and why is it considered token-efficient?

EnGram is a conditional memory architecture from DeepSeek designed to address a transformer limitation: lack of native knowledge lookup. Instead of using multiple attention layers and expensive reasoning tokens to perform tasks that should be simple lookups, EnGram retrieves short sequences (about 2–3 tokens) via hash functions from a large embedding table. It then filters retrieved patterns against the current context using a gating mechanism. The result is substantial performance gains without spending many extra tokens, positioning it as a route toward more factual memory.

How does Kilo Code’s app builder strategy differ from Lovable, Replit, and Cursor?

Kilo Code launched its app builder after a six-week sprint following a late-December release, backed by $8 million in seed funding. Its positioning targets actual engineers rather than non-technical users, with the CEO describing a desire to compete in the space of VS Code—open-source and engineering-friendly—rather than being a “vibe coding” tool for casual users. The transcript frames this as a platform-breadth play inspired by GitLab’s approach: ship a comprehensive toolchain, integrate with existing workflows, and emphasize reliability and flexibility. The open question is whether an engineer-first niche can sustain a fourth major player between Lovable and Cursor.

Review Questions

What evidence suggests investors are willing to fund XAI despite safety failures, and what contracts or metrics are cited?
How do Amodei and Hassabis differ on AGI timing and on why jobs may not be fully automatable?
What mechanism does EnGram add to transformers to improve factual lookup efficiency, and how does it reduce token cost?

Key Points

1
XAI’s $20 billion upsized Series E (about $230B valuation) signals that only a few AI labs have the capital depth to endure multi-year scaling costs.
2
XAI’s funding proceeded despite Grok deepfake incidents and regulatory probes across the EU, UK, India, Malaysia, and France, implying investors expect safety and product maturity over time.
3
Grok’s expansion includes a Department of Defense deal positioning it as the DoD’s AI agents platform, alongside consumer and market integrations like Poly Market and Call.
4
At Davos, Dario Amodei forecast AGI in 2026–2027 via AI-assisted coding loops, while Demis Hassabis estimated a 50% chance by decade’s end and emphasized the “last 5%” of human job skills.
5
Apple’s multi-year Gemini-based foundation model collaboration shifts model leverage toward Google, reportedly costing Apple about $1 billion per year and involving a custom 1.2 trillion-parameter Gemini model.
6
DeepSeek’s EnGram targets transformer “memory” limitations by adding hash-based conditional retrieval with gating, aiming for token-efficient factual lookup.
7
Kilo Code’s six-week app builder launch targets engineers with a VS Code-like, open-source-friendly platform strategy aimed at competing with Lovable, Replit, and Cursor.

Highlights

XAI raised $20 billion even while Grok deepfakes triggered investigations across multiple countries—then secured a Department of Defense deal anyway.

Amodei’s AGI timeline (2026–2027) hinges on AI writing code and humans reviewing, while Hassabis argues jobs persist because humans supply the hardest-to-automate “last 5%.”

Apple’s Gemini-based foundation model partnership reportedly costs $1 billion per year and includes a custom 1.2 trillion-parameter Gemini model, reshaping default AI distribution pressure on OpenAI.

EnGram reframes factual memory as conditional retrieval: short-token hash lookup plus gating to avoid expensive reasoning tokens.

Kilo Code’s engineer-first “VS Code for app building” pitch arrives after a six-week sprint and positions it as a potential fourth player in vibe coding.

Topics

XAI Funding
AGI Timeline
Apple Gemini Partnership
EnGram Memory
Kilo Code App Builder