Issue 16 · one issue a week — the judgment is continuous
JustSayAIWeeklyReport
Kolibri-1 Goes Open Source, Gemini Cuts Free Tier: The Old LLM Era Is Over
人民公园说AI · JustSayAI
www.justsayai.org
Week of
09.29 — 10.05
2026-10-05 · the week's read
Contents · Chapters
This issue,
what to read.
The timeline on the left mirrors these chapters — read in order, or jump straight to any story.
Editor’s Thesis
The Week's Judgment
Kolibri-1 Goes Open Source,Gemini Cuts Free Tier:
The Old LLM Era Is Over
The open-sourcing of Kolibri-1 stands in sharp contrast to Google's tightening of Gemini's free-tier allowances, signaling that the AI industry is moving beyond the old paradigm of monolithic super-models and unconditional free expansion. DeepMind's concept of symbiotic Agents further dismantles the singularity fantasy of all-capable models, as technical approaches begin to diverge. Notably, Google is simultaneously showcasing the engineering value of complex inference through Gemini 3.8 Flash while constraining its free layer—a tension that makes the interplay between commercialization and open-source ecosystems a critical variable in shaping what comes next.
The Tape · This Week's Audio
The full 07:48. Hear the first 60s free.
JustSayAI Podcast · Weekly issue 16 audio
First 60s free · full 07:48 for subscribers
Key Moments · every red tick is a marked moment in this week's audio
00:13in preview
Kolibri Skips the Race for #1, Bets on Lowest Cost Per Token
“Kolibri, 78 billion parameters, only 3.46 billion active—that number just looks off.”
- Kolibri uses sparse activation: 3.46B of 78B parameters active, trading breadth for per-token cost efficiency in common scenarios.
- Apache 2.0 and million-context claims signal giving up universal dominance for cost-per-token competition.
- Industry logic shifted from leaderboard rankings to cost efficiency; Kolibri explicitly avoids chasing first place, taking a fundamentally different business path.
01:50locked · past preview
Artificial Symbiotic Intelligence: From Single-Model Singularity to Multi-Agent Organizational Efficiency
“Okay, second point — that "artificial symbiotic intelligence" thing from DeepMind? The moment I heard the term, something just felt off.”
- Artificial Symbiotic Intelligence isn't mere rhetoric—it acknowledges diminishing marginal returns for single LLMs, where scaling compute 10× may not yield 10× smarter capabilities.
- Five specialized Agents cross-checking and calling each other beats 10x parameters, cheaper too.
- Swapping individual IQ for organizational efficiency: five average people with a good process often beat one Einstein.
04:21locked · past preview
Google Argon: Safety Policy Is Also a Compute Valve
“Right after dropping Gemini Argon, they turned around and slashed the free tier.”
- Argon limits access to cybersecurity citing safety, but high inference costs prevent full opening.
- Safety is Google's cost-control valve, not just protection.
- New model launch cuts free tier, confirming LLM compute cost pressure is real.
Market Tape · The Week's Prices
Prices first, judgment after.
3 AI names · 1 up · 2 downBest GOOGL +0.22%Worst BABA -2.68%2026-09-29 → 2026-10-05 · Closing prices against the last session before the window. Those are facts; what we draw from them is judgment.
GOOGL · week's best
+0.22%
342.75 → 343.5
worst session -1.70%（2026-10-01）
worst session -1.70%（2026-10-01）
688111.SH
-1.44%
222 → 218.8
worst session -2.30%（2026-09-29）
worst session -2.30%（2026-09-29）
BABA · week's worst
-2.68%
108.76 → 105.85
worst session -1.49%（2026-10-02）
worst session -1.49%（2026-10-02）
UpDownWeekly change · source: third-party market data
Shanghai Composite +0.31%（3842.19）Nasdaq Composite +1.19%（27190.86）
Story 01 · Tape 00:13 · AI Industry News
Kolibri Skips the Race for #1, Bets on Lowest Cost Per Token
Kolibri, 78 billion parameters, only 3.46 billion active—that number just looks off.
Fact
🤖 Aleph-Alpha Open-Sources Kolibri-1: 78B Params, 3.46B Active, 1M Token Context
- 01Kolibri uses sparse activation: 3.46B of 78B parameters active, trading breadth for per-token cost efficiency in common scenarios.
- 02Apache 2.0 and million-context claims signal giving up universal dominance for cost-per-token competition.
- 03Industry logic shifted from leaderboard rankings to cost efficiency; Kolibri explicitly avoids chasing first place, taking a fundamentally different business path.
Source ↗ huggingface.co
Story 02 · Tape 01:50 · AI Research
Artificial Symbiotic Intelligence: From Single-Model Singularity to Multi-Agent Organizational Efficiency
Okay, second point — that "artificial symbiotic intelligence" thing from DeepMind? The moment I heard the term, something just felt off.
Fact
🧠 DeepMind's "Artificial Symbiotic Intelligence": General AI May Emerge from Multi-Agent Networks
- 01Artificial Symbiotic Intelligence isn't mere rhetoric—it acknowledges diminishing marginal returns for single LLMs, where scaling compute 10× may not yield 10× smarter capabilities.
- 02Five specialized Agents cross-checking and calling each other beats 10x parameters, cheaper too.
- 03Swapping individual IQ for organizational efficiency: five average people with a good process often beat one Einstein.
Source ↗ the-decoder.com
Story 03 · Tape 04:21 · Technology
Google Argon: Safety Policy Is Also a Compute Valve
Right after dropping Gemini Argon, they turned around and slashed the free tier.
Fact
🤖 Google Unveils Gemini 4 Argon: Safety-First, Cybersecurity Debut
- 01Argon limits access to cybersecurity citing safety, but high inference costs prevent full opening.
- 02Safety is Google's cost-control valve, not just protection.
- 03New model launch cuts free tier, confirming LLM compute cost pressure is real.
Source ↗ nytimes.com
Signal Scan · The Week's Other Signals
2 more that missed the main thread.
🤖 Google slashes Gemini free tier days after Argon launch: only Flash-Lite remains free from Oct 9
This issue05🤖 Google Unveils Gemini 3.8 Flash: 4 Developers Demo Complex inference Apps
This issueThe full week — every daily brief's headline, linked to its issue:
09.29Shopify opens checkout to browser AI agents; AI attacks—AI Daily BriefShopify opens checkout to Agent; Nvidia $150B buyback; Claude Sonnet 5.5 on AWS.
AM09.29Anthropic IPO Loss, Growth; AMD Buys World Labs $8.2B - AI Daily BriefAnthropic IPO: AI resists shutdown; AMD buys World Labs; recursive AI ban.
PM09.30OpenAI Launches Dots Proactive AI Agent, Anthropic Drops GLM-5.3 Benchmark Report - AI Daily Brief (Sep 30)OpenAI drops proactive Agent Dots and budget LLM GPT-6.1 Sol as China's generative AI users hit 700M, while Trump's light-touch regulation clashes with industry safety warnings—AI is penetrating society at unprecedented speed.
AM09.30OpenAI in $30B Bridge Loan Talks, AI Resurrects Greta Garbo for Bearing Ad | AI Daily Brief
PM10.01🚀 Trump AI Pledge Typo; BoE Gov Urges AI Intervention - AI Daily BriefTrump AI pledge typo; Bank of England, Google Gemini 4 Argon tighten AI safety.
AM10.01Trump Tweaks America.gov on Politics, Anthropic's IPO Pitch: Stronger Models, Safer AI Paradox - AI Daily Brief (Oct 1, 3PM)Trump team hardcodes political blocks into AI bots; Anthropic's IPO pitch touts stronger models as safer—paradox alert; DeepSeek-Huawei team on Ascend tools. Political meddling, safety theater, chip wars reshape AI.
PM10.02Armadin Raises $255.5M Series B, Study Exposes Gender Bias in Claude & GPT | AI Daily BriefArmadin raises $250M at $2.5B+ valuation as enterprise AI launches Agent security, virtual human interaction & open decision models—while Claude & GPT's latest versions reveal severe gender bias.
AM10.02Grok Accused of Pushing Trump to Capture Venezuela Leader, LeCun Dismisses AI Extinction Fears as Amodei's "Delusion" - JustSayAI AI Daily Brief (Oct 2, 3PM)Musk's Grok urged Trump to arrest foreign presidents, AI godfather LeCun dismisses extinction fears, while NVIDIA, OpenAI, Microsoft and Suno unleash voice and ultra-fast LLMs—tech sprint collides with political ethics firestorm.
PM10.03GPT-6 Astra Cracks Napoleonic War Letters, Georgia Holds Emergency Meeting on AI-Linked Secret Ballots | AI Daily BriefGPT-6 Astra cracks 217-year-old Napoleonic sealed letter in 6 hours; AI advances in blind wargaming & music generation, yet election privacy flaws and hospital algorithm scheduling controversies emerge.
AM10.03DeepSeek & Huawei Break Free From CUDA, Apple Locks Down macOS Full Disk Access | AI Daily Brief (Oct 3, 3PM)DeepSeek-Huawei open-source toolkit targets Nvidia ecosystem lock-in; Apple tightens iOS privacy to curb Agent abuse; industry debates why Agents code well but fail at business logic.
PM10.04Google Kills Dual Gemini Tiers, NVIDIA Shield TV Pro Jumps $100 - AI Daily Brief (Oct 4, 5AM)Google abruptly cuts Gemini model access for free and Plus users alike; Nvidia hikes Shield TV prices against the tide; Claude Opus 5.5 dominates dev mindshare with top-tier coding; AI liability debates ignite.
AM10.04Google Gemini Argon Cuts Free Tier After Launch, Arizona Court Overturns Verdict Over AI Victim Video | AI Daily BriefGoogle slashes Gemini Argon free tier as US pushes AI globally, while Arizona court overturns murder conviction over AI video evidence—a three-way tension exposing the razor's edge between AI commercialization and judicial ethics.
PMThis week ran 5 headlines; 3 made the main thread; 12 daily briefs.
The Archive · Past Issues
One issue a week. The judgment is continuous.
2026.10.05Kolibri-1 Goes Open Source, Gemini Cuts Free Tier: The Old LLM Era Is Over●2026.09.28Monetization, Agent Security Pressure: DeepSeek V4-Pro Peak Pricing→2026.09.21Sovereign AI Warfare: Scaleout Deploys Autonomous Drone Strike Models→2026.09.14DeepMind Co-Founder Backs Amodei: Benchmark Arms Race Forces AI Safety Consensus→2026.09.07AI Reshapes Creation and Jobs: Google Launches Lyria 3.5→
The Week in One Line · One line
“Kolibri-1 Goes Open Source, Gemini Cuts Free Tier: The Old LLM Era Is Over”
Two issues a day. Ten minutes to turn AI noise into judgment — mornings for the world, evenings for China.
人民公园说AIJustSayAI Weekly Report
JustSayAIWeeklyReport
Two issues a day — AI noise into judgment.
Next Issue · next
2026.10.12 · every Monday →